Abstract
We show that PRAMs can be obliviously simulated with perfect security, incurring only \(O(\log N \log \log N)\) blowup in parallel runtime, \(O(\log ^3 N)\) blowup in total work, and O(1) blowup in space relative to the original PRAM. Our results advance the theoretical understanding of Oblivious (Parallel) RAM in several respects. First, prior to our work, no perfectly secure Oblivious Parallel RAM (OPRAM) construction was known; and we are the first in this respect. Second, even for the sequential special case of our algorithm (i.e., perfectly secure ORAM), we not only achieve logarithmic improvement in terms of space consumption relative to the stateoftheart, but also significantly simplify perfectly secure ORAM constructions. Third, our perfectly secure OPRAM scheme matches the parallel runtime of earlier statistically secure schemes with negligible failure probability. Since we remove the dependence (in performance) on the security parameter, our perfectly secure OPRAM scheme in fact asymptotically outperforms known statistically secure ones if (sub)exponentially small failure probability is desired. Our techniques for achieving small parallel runtime are novel and we employ special expander graphs to derandomize earlier statistically secure OPRAM techniques—this is the first time such techniques are used in the constructions of ORAMs/OPRAMs.
An online full version of our paper [9] is available at https://eprint.iacr.org/2018/364.
Download conference paper PDF
1 Introduction
Oblivious RAM (ORAM), originally proposed in the groundbreaking work by Goldreich and Ostrovsky [21, 22], is an algorithmic technique that transforms any RAM program to a secure version, such that an adversary learns nothing about the secret inputs from observing the program’s access patterns to memory. The parallel extension of ORAM was first phrased by Boyle, Chung, and Pass [6]. Similar to ORAM, an Oblivious Parallel RAM (OPRAM) compiler transforms a Parallel RAM (PRAM) program into a secure form such that the resulting PRAM’s access patterns leak no information about secret inputs. ORAMs and OPRAMs have been recognized as powerful building blocks in both theoretical applications such as multiparty computation [5, 25, 29], as well as in practical applications such as cloud outsourcing [14, 37, 40], and secure processor design [17, 18, 28, 30, 31, 35].
Henceforth in this paper, we consider ORAMs to be a special case of OPRAMs, i.e., when both the original PRAM and the OPRAM have only one CPU. To characterize an OPRAM scheme’s overhead, we will use the standard terminology total work blowup to mean the multiplicative increase in total computation comparing the OPRAM and the original PRAM; and we use the term depth blowup to mean the multiplicative increase in parallel runtime comparing the OPRAM and the original PRAM—assuming that the OPRAM may employ more CPUs than the original PRAM to help parallelize its computation [7]. Note that for the case of sequential ORAMs, total work blowup is equivalent to the standard notion of simulation overhead [21, 22], i.e., the multiplicative increase in runtime comparing the ORAM and the original RAM. Finally, we use the term space blowup to mean the multiplicative blowup in space when comparing the OPRAM (or ORAM) and that of the original PRAM (or RAM).
The original ORAM schemes, proposed by Goldreich and Ostrovsky [21, 22], achieved polylogarithmic overheads but required the usage of pseudorandom functions (PRFs); thus they defend only against computationally bounded adversaries. Various subsequent works [2, 10, 12, 13, 36, 38, 39], starting from Ajtai [2] and Damgård et al. [13] investigated informationtheoretically secure ORAM/OPRAM schemes, i.e., schemes that do not rely on computational assumptions and defend against even unbounded adversaries. As earlier works point out [2, 13], the existence of efficient ORAM schemes without computational assumptions is not only theoretically intriguing, but also has various applications in cryptography. For example, informationtheoretically secure ORAM schemes can be applied to the construction of efficient RAMmodel, informationtheoretically secure multiparty computation (MPC) protocols [4]. Among known informationtheoretically secure ORAM/OPRAM schemes [2, 6, 10,11,12,13, 36, 38, 39], almost all of them achieve only statistical security [2, 6, 10,11,12, 36, 38, 39], i.e., there is still some nonzero failure probability—either correctness or security failure—but the failure probability can be made negligibly small in N where N is the RAM/PRAM’s memory size. Damgård et al. [13] came up with the first perfectly secure ORAM construction—they achieve zero failure probability against computationally unbounded adversaries. Although recent works have constructed statistically secure OPRAMs [6, 10, 11], there is no known (nontrivial) perfectly secure OPRAM scheme to date.
Motivation for Perfect Security. Perfectly secure ORAMs/OPRAMs are theoretically intriguing for various reasons:

1.
First, to achieve \(2^{\kappa }\) failure probability (either in security or in correctness), the best known statistically secure OPRAM scheme [7, 10] incurs a \(O(\kappa \log N)\) total work blowup and \(O(\log \kappa \log N)\) depth blowup where N is the PRAM’s memory size. Although for negligibly small in N failure probability the blowups are only polylogarithmic in N, they can be as large as \(N^c\) for some constant \(c < 1\) if one desires (sub)exponentially small failure probability in N.

2.
Second, perfectly secure ORAM schemes have been used as a building block in recent results on oblivious algorithms [3, 36] and searchable encryption schemes [15]. Typically these algorithmic constructions rely on divideandconquer to break down a problem into smaller sizes and then apply ORAM to a small instance—since the instance size N is small (e.g., logarithmic in the security parameter), negligible in N failure probability is not sufficient and thus these works demand perfectly secure ORAMs/OPRAMs and existing statistically secure schemes result in asymptotically poorer performance.

3.
Third, understanding the boundary of perfect and statistical security has been an important theoretical question in cryptography. For example, a longstanding open problem in cryptography is to separate the classes of languages that admit perfect ZK and statistical ZK proofs. For ORAMs/OPRAMs too, it remains open whether there are any separations between statistical and perfect security (and we believe that this is an exciting future direction). Perfect security is also useful in other contexts such as multiparty computation (MPC). For example, Ishai et al. [26] and Genkin et al. [19] show that perfectly secure MPC is required to achieve their respective goals matching the “circuit complexity” of the underlying application. Perfectly secure ORAMs/OPRAMs can enable perfectly secure RAMmodel MPC, and thus we believe that they can be an important building block in other areas of theoretical cryptography.
1.1 Our Results and Contributions
In this paper, we prove the following result which significantly advances our theoretical understanding of perfectly secure ORAMs and OPRAMs in multiple respects. We present the informal theorem statement below and then discuss its theoretical significance.
Theorem 1
(Informal statement of main theorem). Any PRAM that consumes N memory blocks each of which is at least \(\log N\)bits long^{Footnote 1} can be simulated by a perfectly oblivious PRAM, incurring \(O(\log ^3 N)\) total work blowup, \(O(\log N \log \log N)\) depth blowup, and O(1) space blowup.
The above theorem improves the theoretical state of the art on perfectly secure ORAMs/OPRAMs in multiple dimensions:

1.
First, our work gives rise to the first perfectly secure (nontrivial) OPRAM construction. No such construction was known before and it is not clear how to directly parallelize the perfectly secure ORAM scheme by Damgård et al. [13].

2.
Second, even for the sequential special case, we improve Damgård et al. [13] asymptotically by reducing a \(\log N\) factor in the ORAM’s space consumption.

3.
Third, our perfectly secure OPRAM’s parallel runtime matches the best known statistically secure construction [7, 10] for negligibly small in N failure probabilities;

4.
Finally, when (sub)exponentially small (in N) failure probabilities are required, our perfectly secure OPRAM scheme asymptotically outperforms all known statistically secure constructions both in terms of total work blowup and depth blowup. For example, suppose that we require \(2^{\kappa }\) failure probability and \(N = \mathsf{poly} (\kappa )\)—then all known statistically secure OPRAM constructions [6, 10, 11] would incur at least \(N^{c}\) total work blowup and \(\varOmega (\log ^2 N)\) depth blowup and thus our new perfectly secure OPRAM construction is asymptotically better for this scenario.
Theorem 1 applies to general block sizes. We additionally show that for sufficiently large block sizes, there exists a perfectly secure OPRAM construction with \(O(\log ^2 N)\) total work blowup and \(O(\log m + \log \log N)\) depth blowup where m denotes the number of CPUs of the original PRAM. Finally, we point out that this work focuses mostly on the theoretical understanding of perfect security in ORAMs/OPRAMs, and we leave it as a future research direction to investigate their practical performance (see also Sect. 6).
Technical Highlights. Our most novel and nontrivial technical contribution is the use of expander graphs techniques, allowing our OPRAM to achieve as small as \(O(\log N \log \log N)\) depth blowup. To the best of our knowledge, this is the first time such techniques have been used in the construction of ORAM/OPRAM schemes. Besides this novel technique, our scheme requires carefully weaving together many algorithmic tricks that have been used in earlier works [7, 10, 21, 22].
1.2 Related Work
Oblivious RAM (ORAM) was first proposed in a groundbreaking work by Goldreich and Ostrovsky [21, 22]. Goldreich and Ostrovsky first showed a computationally secure ORAM scheme with polylogarithmic simulation overhead. Therefore, one interesting question is whether ORAMs can be constructed without relying on computational assumptions. Ajtai [2] answered this question and showed that statistically secure ORAMs with polylogarithmic simulation overhead exist. Although Ajtai removed computational assumptions from ORAMs, his construction has a (negligibly small) statistical failure probability, i.e., with some negligibly small probability, the ORAM construction can leak information. Subsequently, Shi et al. [36] proposed the treebased paradigm for constructing statistically secure ORAMs. Treebased constructions were later improved further in several works [10, 12, 20, 38, 39], and this line of works improve the practical performance of ORAM by several orders of magnitude in comparison with earlier constructions. It was also later understood that the treebased paradigm can be used to construct computationally secure ORAMs saving yet another \(\log \log \) factor in cost in comparison with statistical security [10, 16].
Perfect security requires that the (oblivious) program’s memory access patterns be identically distributed regardless of the inputs to the program; and thus with probability 1, no information can be leaked about the secret inputs to the program. Perfectly secure ORAM was first studied by Damgård et al. [13]. Their construction achieves \(O(\log ^3 N)\) simulation overhead and \(O(\log N)\) space blowup relative to the original RAM program. Their construction is a Las Vegas algorithm and there is a negligibly small failure probability that the algorithm exceeds the stated runtime. Raskin et al. [34] and Demertzis et al. [15] achieve a worstcase bandwidth of \(O(\sqrt{N}\frac{\log N}{\log \log N})\) and \(O(N^{1/3})\), respectively. As mentioned, even for the sequential case, our paper asymptotically improves Damgård et al.’s result [13] by avoiding the \(O(\log N)\) blowup in space; and moreover, our ORAM construction is conceptually simpler than that of Damgård et al.’s.
Oblivious Parallel ORAM (OPRAM) was first proposed in an elegant work by Boyle et al. [6], and subsequently improved in several followup works [7, 8, 10, 11, 32]. All known results on OPRAM focus on the statistically secure or the computationally secure setting. To the best of our knowledge, until this paper, we know of no efficient OPRAM scheme that is perfectly secure. Chen et al. [11] introduced a generic method to transform any ORAM into an OPRAM at the cost of a \(\log N\) blowup. Their techniques achieve statistical security since security (or correctness) is only guaranteed with high probability (specifically, when some queue does not become overloaded in their scheme).
Defining a good performance metric for OPRAMs turned out to be more interesting and nontrivial than for ORAMs. Boyle et al. [6] were the first to define a notion of simulation overhead for OPRAM: if an OPRAM’s simulation overhead is X, it means that if the original PRAM consumes m CPUs and completes in parallel runtime T, then the oblivious counterpart must complete within \(X \cdot T\) time also consuming m CPUs. The recent work of Chan et al. [7] observes that if the OPRAM could consume more CPUs than the original PRAM, then the oblivious simulation can benefit from the additional parallelism and be additionally sped up by asymptotic factors. Under the assumption that the OPRAM can consume more CPUs than the original PRAM, Chan et al. [7, 10] show that statistically secure OPRAM schemes can be constructed with \(O(\log ^2 N)\) blowup in total work and only \(\widetilde{O}(\log N)\) blowup in depth (where depth characterizes the parallel runtime of a program assuming ample number of CPUs). Our paper is the first to construct an OPRAM scheme with perfect security, and our OPRAM’s depth matches existing schemes with statistical security assuming negligible in N security failure; however, if (sub)exponentially small failure probability is required, our new OPRAM scheme can asymptotically outperform all known statistically secure OPRAMs!
2 Technical Roadmap
In this section, we present an informal roadmap of our technical approach to aid understanding.
2.1 Simplified Perfectly Secure ORAM with Asymptotically Smaller Space
First, we propose a perfectly secure ORAM scheme that is conceptually simpler than that of Damgård et al. [13] and gains a logarithmic factor in space. Our construction is inspired by the hierarchical ORAM paradigm originally proposed by Goldreich and Ostrovsky [21, 22]. However, most existing hierarchical ORAMs achieve only computational security since they rely on a pseudorandom function (PRF) for looking up hash tables in the hierarchical data structure. Thus our focus is to get rid of this PRF and achieve perfect security.
Background: Hierarchical ORAM. The recent work by Chan et al. [8] gave a clean and modular exposition of the hierarchical paradigm. A hierarchical ORAM consists of \(O(\log N)\) levels that are geometrically increasing in size. Specifically, level i is capable of storing \(2^i\) memory blocks. One could think of this hierarchical data structure as a hierarchy of stashes where smaller levels act as stashes for larger levels. In existing schemes with computational security, each level is an oblivious hashtable [8]. To access a block at logical address \({\mathsf {addr}} \), the CPU sequentially looks up every level of the hierarchy (from small to large) for the logical address \({\mathsf {addr}} \). The physical location of a logical address \({\mathsf {addr}} \) within the oblivious hashtable is determined using a PRF whose secret key is known only to the CPU but not to the adversary. Once the block has already been found in some level, for all subsequent levels the CPU would just look for a dummy element, denoted by \(\bot \). When a requested block has been found, it is marked as deleted in the corresponding level where it is found. Every \(2^i\) memory requests, we perform a rebuild operation and merge all levels smaller than i (including the block just fetched and possibly updated if this is a write request) into level i—at this moment, the oblivious hashtable in level i is rebuilt, where every block’s location in the hash table is determined using a PRF.
As Chan et al. [8] point out, the hierarchical ORAM paradigm effectively reduces the problem of constructing ORAM to constructing an oblivious hashtable supporting two operations: (1) rebuild takes in a set of blocks each tagged with its logical address, and constructs a hashtable data structure that facilitates lookups later; and (2) lookup takes a request that is either a logical address \({\mathsf {addr}} \) or dummy (denoted \(\bot \)), and returns the corresponding block requested. Obliviousness (defined w.r.t. the joint access patterns of the rebuild and lookup phases) is guaranteed as long as during the lifetime of the oblivious hashtable, the sequence of lookup requests never ask for the same real element twice—and this invariant is guaranteed by the specific way the hierarchical ORAM framework uses the oblivious hashtable as a building block (more specifically, the fact that once a block is found, it is moved to a smaller level and a dummy block is requested from all subsequent levels).
Removing the PRF. As mentioned, an oblivious hashtable relies on a PRF to determine each block’s location within a hashtable instance; and both the rebuilding phase and the lookup phase use the same PRF for placing and fetching blocks respectively. Since we wish to achieve perfect security, we would like to remove the PRF. One simple idea is to randomly permute all blocks within a level— this way, each lookup of a real block would visit a random location and we could hope to retain security as long as every real block is requested at most once for every level (in between rebuilds)^{Footnote 2}. Using techniques from earlier works [7, 10], it is possible to obliviously perform such a random permutation without disclosing the permutation; however, difficulty arises when one wishes to perform a look up—if blocks are randomly permuted within a level during rebuild, lookup must know where each block resides to proceed successfully. Thus if the CPU could hold a position map for free to remember where each block is in the hierarchical data structure, the problem would have been resolved: during every lookup, the CPU could first look up the physical location of the logical address requested, and then proceed accordingly.
Actually storing such a position map, however, would consume too much CPU space. To avoid storing this position map, we are inspired by the recursion technique that is commonly adopted by treebased ORAM schemes [36]—however, as we point out soon, making the recursion idea work for the hierarchical ORAM paradigm is more sophisticated. The highlevel idea is to recursively store the position map in a smaller ORAM rather than storing it on the CPU side; we could then recurse and store the position map of the position map in an even smaller ORAM, and so on—until the ORAM’s size becomes O(1) at which point we would have the CPU store the entire ORAM. Henceforth, we use the notation \(\mathsf{ORAM}_D\) to denote the ORAM that stores the actual data blocks where \(D = O(\log N)\); and we use \(\mathsf{ORAM}_d\) to denote the ORAM at depth d of this recursion where \(d \in [0..D1]\). Thus, the larger d is, the larger the ORAM.
Although this recursion idea was very simple in the treebased paradigm, it is not immediately clear how to make the same recursion idea work in the hierarchical ORAM paradigm. One trickiness arises since in a hierarchical ORAM, every \(2^i\) requests, the ORAM would reshuffle and merge all levels smaller than i into level i — this is called a rebuild of level i. When a leveli rebuild happens, the position labels in the positionmap ORAM must be updated as well to reflect the blocks’ new locations. In a similar fashion, the position labels in all of \(\mathsf{ORAM}_0, \mathsf{ORAM}_1, \ldots , \mathsf{ORAM}_{D1}\) must be updated. We make the following crucial observation that will enable a coordinated rebuild technique which we will shortly explain:
(Invariant necessary for coordinated rebuild:) If a data block resides at level i of \(\mathsf{ORAM}_D\), then its position labels in all recursion depths must reside in level i or smaller^{Footnote 3}.
This invariant enables a coordinated rebuild technique: when the data ORAM (i.e., \(\mathsf{ORAM}_D\)) merges all levels smaller than i into level i, all smaller recursion depths would do the same (unless the recursion depth is too small and does not have level i, in which case the entire ORAM would be rebuilt). During this coordinated rebuild, \(\mathsf{ORAM}_D\) would first perform its rebuild, and propagate the position labels of all blocks involved in the rebuild to recursion depth \(D1\); then \(\mathsf{ORAM}_{D1}\) would perform its rebuild based on the position labels learned from \(\mathsf{ORAM}_D\), and propagate the new position labels involved to recursion depth \(D2\), and so on. As we shall discuss in the technical sections, rebuilding a level (in any recursion depth) can be accomplished through the help of O(1) oblivious sorts and an oblivious random permutation.
Handling Dummy Blocks with Oblivious Linked Lists. The above idea almost works, but not quite so. There is an additional technical subtlety regarding how to handle and use dummy blocks. Recall that during a memory access, if a block requested actually resides in a hierarchical level, we would read the memory location that contains the block (and this memory location could be retrieved through a special recursive position map technique). If a block does not reside in a level (or has been found in a smaller level), we still need to read a dummy location within the level to hide the fact that the block does not reside within the current level.
Recall that the ith level must support up to \(2^i\) lookups before the level is rebuilt. Thus, one idea is to introduce \(2^i\) dummy blocks, and obliviously and randomly permute all blocks, real and dummy alike, during the rebuild. All dummy blocks may be indexed by a dummy counter, and every time one needs to look up a dummy block in a level, we will visit a new dummy block. In this way, we can retain obliviousness by making sure that every real block and every dummy block is visited at most once before the level is rebuilt again.
To make this idea fully work, there must be a mechanism for finding out where the next dummy block is every time a dummy lookup must be performed. One naïve idea would be to use the same recursion technique to store position maps for dummy blocks too—however, since each memory request might involve reading \(O(\log N)\) dummy blocks, one per level, doing so would incur extra blowup in runtime and space. Instead, we use an oblivious dummy linked list to resolve this problem—this oblivious dummy linked list is inspired by technical ideas in the Damgård et al. construction [13]. In essence, each dummy block stores the pointer to the next dummy block, and the head pointer of the linked list is stored at a designated memory location and updated upon each read of the linked list. In the subsequent technical sections, we will describe how to rely on oblivious sorting to rebuild such an oblivious dummy linked list to support dummy lookups.
Putting It Altogether. Putting all the above ideas together, the formal presentation of our perfectly secure ORAM scheme adopts a modular approach^{Footnote 4}. First, we define and construct an abstraction called an “oblivious onetime memory”. An oblivious onetime memory allows one to obliviously create a data structure given a list of input blocks. Once created, one can look up real or dummy blocks in the data structure, and to look up a real block one must provide a correct position label indicating where the block resides (imagine for now that the position label comes from an “oracle” but in the full ORAM scheme the position label comes from the recursion). An oblivious onetime memory retains obliviousness as long as every real block is looked up at most once and moreover, dummy blocks are looked up at most n times where n is a predetermined parameter (that the scheme is parametrized with).
Once we have this “oblivious onetime memory” abstraction, we show how to use it to construct an intermediate abstraction referred to as a “positionbased ORAM”. A positionbased ORAM contains a hierarchy of oblivious onetime memory instances, of geometrically growing sizes. A positionbased ORAM is almost a fully functional ORAM except that we assume that upon every memory request, an “oracle” will somehow provide a correct position label indicating where the requested block resides in the hierarchy.
Finally, we go from such a “positionbased ORAM” to a fully functional ORAM using the special recursive positionmap technique as explained. At this point, we have constructed a perfectly secure ORAM scheme with \(O(\log ^3 N)\) simulation overhead. Specifically, one \(\log N\) factor arises from the \(\log N\) depths of recursion, the remaining \(\log ^2 N\) factor arises from the cost of the ORAM at each recursion depth. Intuitively, our perfectly secure ORAM is a logarithmic factor more expensive than existing computationallysecure counterparts in the hierarchical framework [8, 23, 27] since the computationallysecure schemes [8, 23, 27] avoid the recursion by adopting a PRF to compute the pseudorandom position labels of blocks.
2.2 Making Our ORAM Scheme Parallel
Our next goal is to make our ORAM scheme parallel. Instead of compiling a sequential RAM program to a sequential ORAM, we are now interested in compiling a PRAM program to an OPRAM. In this section, we describe an informal roadmap of our technical approach to parallelism. However, due to lack of space, we defer the details to the full version of our paper [9].
When the OPRAM Consumes the Same Number of CPUs as the PRAM. Suppose that the original program is a PRAM that completes in T parallel steps consuming m CPUs. We now would like to parallelize our earlier ORAM scheme and construct an OPRAM that completes in \(T \cdot O(\log ^3 N)\) parallel steps consuming also exactly m CPUs. To accomplish this, first, we need to parallelize within each positionbased ORAM so m CPUs can perform work concurrently. This is not too difficult to accomplish given the simplicity of our positionbased ORAM construction. Next, when m CPUs have all fetched position labels at one recursion depth, they need to pass these position labels to the CPUs at the next depth. The main technique needed here is oblivious routing: when the m CPUs at recursion depth d have fetched the position labels for the next recursion depth, the m CPUs at depth d must now obliviously route the position labels to the correct fetch CPU at the next recursion depth. As shown in earlier works [6, 7, 10], such oblivious routing can be accomplished with m CPUs in \(O(\log m)\) parallel steps.
We stress that the simplicity of our sequential ORAM construction makes it easy to parallelize — in comparison, we are not aware how to parallelize Damgård et al. [13]’s construction^{Footnote 5}.
When the OPRAM May Consume Unbounded Number Of CPUs. The more interesting question is the following: if the OPRAM is allowed to consume more CPUs than the original PRAM, can we further reduce its parallel runtime? If so, it intuitively means that the overheads arising due to obliviousness are parallelizable in nature. This model was first considered by Chan et al. [7] and can be considered as a generalization of the case when the OPRAM must consume the same number of CPUs as the original PRAM.
So far, in our OPRAM scheme, although within each recursion depth, up to m requests can be served concurrently, the operations over all \(O(\log N)\) recursion depths must be performed sequentially. There are two reasons that necessitate this sequentiality:

1.
Fetch phase: first, to fetch from recursion depth d, one must wait for the appropriate position labels to be fetched from recursion depth \(d1\) and routed to recursion depth d;

2.
Maintain phase: recall that coordinated rebuilding (see Sect. 2.1) is performed across all recursion depths in the reverse direction: recursion depth d must rebuild first and then propagate the new positions labels back to recursion depth \(d1\) before \(d1\) can rebuild (recall that recursion depth \(d1\) must store the position labels for blocks in depth d).
Note that for the fetch phase, oblivious routing between any two adjacent recursion depths would consume \(O(\log m)\) depth; for the maintain phase, rebuilding a hierarchical level can consume up to \(O(\log N)\) depth (due to oblivious sorting of up to O(N) blocks). Thus, the current OPRAM algorithm incurs a depth blowup of \(O(\log ^2 N)\) for moderate sizes of m, e.g., when \(\log m = \varTheta (\log N)\). Our next goal is to reduce the depth blowup to \(\widetilde{O}(\log N)\), and this turns out to be highly nontrivial.
Reducing the Depth of the Fetch Phase with Expander Graphs. Using the recursion technique, it seems inherent that one must fetch from smaller recursion depths before embarking on larger ones. To reduce the depth of the fetch phase, we ask whether the depth incurred by oblivious routing in between adjacent recursion depths can be reduced. In the statistically and computationally secure settings, the recent work by Chan, Chung, and Shi have tried to tackle a similar problem for treebased OPRAMs [7]. Their idea is to construct an offline/online routing algorithm. Although the offline phase incurs \(O(\log N)\) depth per recursion depth, the offline work of all recursion depths can be performed concurrently rather than sequentially. On the other hand, the online phase of their routing algorithm must be performed sequentially among the recursion depths, but happily the online routing phase incurs only O(1) depth per recursion depth. Unfortunately, the offline/online routing algorithm of Chan et al. [7] is a randomized algorithm that leverages some form of statistical “load balancing”, and such load balancing can fail with negligibly small probability—this makes their algorithm unsuitable for the perfect security setting.
We propose a novel offline/online routing algorithm that achieves perfect security using special expander graphs—our techniques can be viewed as a method for derandomizing a new variant of the offline/online routing techniques described by Chan et al. [7]. Like Chan et al. [7], our offline/online routing algorithm achieves \(O(\log N)\) depth for each recursion depth in the offline stage but the work in all recursion depths can be performed in parallel in the offline stage. By contrast, the online phase must traverse the recursion depths sequentially, but the online stage of routing can be accomplished in O(1) depth per recursion depth. To achieve this, we rely on a core building block called a “loose compactor”. Leveraging special expander graphs, we show how to build a loose compactor with small online depth—since this part of our techniques are novel, we present a more expanded overview in Sect. 2.3 while deferring a detailed, formal description to the full version [9].
Reducing the Depth of the Maintain Phase. We also must reduce the depth of the maintain phase. Although a naïve implementation of coordinated rebuild is to do it sequentially from recursion depth D down to recursion depth 0, we devise a method for performing the coordinated rebuild in parallel among all recursion depths. Recall that in the naïve solution, recursion depth \(d1\) must wait for recursion depth d to relocate its blocks and be informed of the new position labels chosen before it starts reshuffling.
In our new algorithm, we introduce the notion of a rehearsal step called “mock shuffle” which determines the new positions of each of the blocks. Note that during this step, the newly chosen block contents (position labels) at the recursion depths are not available. Now, instead of sequentially performing the shuffle, in a mock shuffle, every recursion depth performs eager reshuffling without having updated the block’s contents (recall that each block in recursion depth d is supposed to store position labels for the next recusion depth \(d+1\)). After this mock shuffle, all blocks’ new positions are determined though their contents are not known. Each mock reshuffle incurs \(O(\log N)\) depth, but they are independent and can be performed in parallel. At this moment, recursion depth d informs the newly chosen position labels to recursion depth \(d1\)—now recursion depth \(d1\) relies on oblivious routing to deliver each block’s contents to the block. Note that recursion depth \(d1\) has already chosen each block’s position at this point and thus in this content update step, each block’s contents will be routed to the corresponding block and all blocks will maintain their chosen positions.
Using this idea, although each recursion depth incurs \(O(\log N)\) depth for the maintain phase, all recursion depths can now perform the maintainphase operations in parallel.
Additional Techniques. Besides the above, additional tricks are needed to achieve \(\widetilde{O}(\log N)\) depth. For example, within each recursion depth, all the hierarchical levels must be read in parallel during the fetch phase rather than sequentially like in existing hierarchical ORAMs [21, 22], and the result of these fetches can be aggregated using an oblivious select operation incurring \(O(\log \log N)\) depth. It is possible for us to read all the hierarchical levels in parallel since each recursion depth must have received the position labels of all real blocks requested before its fetch phase starts—and thus we know for each requested block which level to look for a real element and which level to visit dummies. We defer additional algorithmic details to the full version [9].
2.3 Offline/Online Routing with Special Expander Graphs
Informal Problem Statement. Without going into excessive details, consider the following abstract problem: imagine that m CPUs at a parent depth have fetched m real or dummy blocks, and each real block contains two position labels for the next depth—thus in total up to 2m position labels have been fetched. Meanwhile, m CPUs at the next depth are waiting to receive m position labels before they can start their fetch. Our task is to obliviously route the (up to) 2m position labels at the parent depth to the m CPUs at the child depth. Using oblivious routing directly would incur \(\varOmega (\log m)\) depth and thus is too expensive.
A Blueprint: Using an Offline/Online Algorithm. As mentioned earlier, our highlevel idea is to leverage an offlineonline paradigm such that the online phase, which must be performed sequentially for all recursion depths, should have small parallel runtime for each recursion depth.
Here is another idea: suppose that we are somehow able to compress the 2m position labels down to m, removing the ones that are not needed by the next recursion depth—this is in fact nontrivial but for now, suppose that somehow it can be accomplished.
Our plan is then the following: in the offline phase, we obliviously and randomly permute the m position labels to be routed (without leaking the permutation), and we obliviously compute the routing permutation \(\pi \) preserving the following invariant: the CPU at position \(\pi (i)\) (in the child depth) is waiting for the ith position label in the permuted array. In other words, the ith position label wants to be routed to the CPU in position \(\pi (i)\); and in the offline phase, we want to route down this \(\pi \).
If we can accomplish all of the above, then in the online phase we simply apply the routing permutation that has been recorded and it takes a single parallel step to complete the routing. Moreover, for the offline phase, as long as we can perform the operations in parallel across all recursion depths, we are allowed to incur \(\log m\) depth.
Informally, obliviousness holds because of the following: recall that the m labels to be routed have been obliviously and randomly permuted. Now, although the routing permutation \(\pi \) is revealed in the online phase, the revealed permutation is uniform at random to an observer.
Technical Challenges: Compaction (and More). The above blueprint seems promising, but there are multiple technical challenges. One critical ingredient that is missing is how to perform compaction from 2m elements down to m, removing the labels not needed by the next recursion depth—in fact, even if we can solve this compaction problem, additional challenges remain in putting these techniques together. However, for the time being, let us focus on the compaction problem alone. The most naïve method is again to leverage oblivious sorting but unfortunately that takes \(\varOmega (\log m)\) depth and thus is too expensive for our purpose.
Pippenger’s FactoryFacility Problem. Our approach is inspired by the techniques described by Pippenger in constructing a selfrouting superconcentrator [33]. Pippenger’s elegant construction can be used to solve a “factoryfacility” problem described as follows. Suppose that 2m factories and m facilities form a special bipartite expander graph: each factory is connected to \({\mathfrak {d}} \) facilities and each facility is connected to \(2{\mathfrak {d}} \) factories, where \({\mathfrak {d}} \) is a constant. Among the factories, m / 64 of them are productive and actually end up manufacturing products. Each productive factory produces \({\mathfrak {d}}/2\) products; these products must be routed to a facility to be stored, and each facility has a storage capacity of \({\mathfrak {d}}/2\). Now, the question is: given the set of productive factories (and assuming that the bipartite graph is known), can we find a satisfying assignment for routing products to facilities, such that (1) every edge in the bipartite graph routes carries at most one unit of flow; (2) all products manufactured are routed; and (3) no facility exceeds its storage capacity.
In his ingenious work [33], Pippenger described a distributed protocol for finding such an assignment: imagine that the factories and facilities are Interactive Turing Machines. Now the factories and facilities exchange messages over edges in the bipartite graph. Pippenger’s protocol completes after \(O(\log m)\) rounds of interaction and a total of O(m) number of messages. Pippenger proved that as long as the underlying bipartite graph satisfies certain expansion properties, his protocol is guaranteed to find a satisfying assignment.
Using Pippenger’s Protocol for Oblivious Loose Compaction. Now we can reduce the problem of (loose) compaction to Pippenger’s factoryfacility problem. Imagine that there are twice as many factories as there are facilities. Another way to think of the factoryfacility problem is the following: imagine that the factories initially store real elements (i.e., the manufactured products) as well as dummies, and in total \(2m \cdot ({\mathfrak {d}}/2)\) amount of storage is consumed since each factory can produce at most \({\mathfrak {d}}/2\) products. We ensure that only m / 64 factories are productive by appropriately adding a constant factor of dummy elements (i.e., dummy factories and facilities). Now, when routed to the facilities, the storage amount is compressed down by a factor of 2 since each facility can store up to \({\mathfrak {d}}/2\) products and the number of facilities is half that of factories. Further, for any satisfying assignment, we guarantee that no real element is lost during the routing, and that is why the compaction algorithm satisfies correctness. Note that such compaction is loose, i.e., we do not completely remove dummies during compaction although we do cut down total storage by a half while preserving all real elements. In our OPRAM algorithm, it turns out that such loose compaction is sufficient, since CPUs who have received dummy position labels can always perform dummy fetch operations.
Pippenger’s protocol can be easily simulated on a PRAM incurring O(m) total work and \(O(\log m)\) parallel runtime—however, a straightforward PRAM simulation of their protocol is not oblivious. In particular, the communication patterns between the factories and facilities (which translate to memory access patterns when simulated on a PRAM) leak information about which factories are productive. Thus it remains for us to show how to obliviously simulate his protocol on a PRAM. We show that this can be done incurring \(O(m \log m)\) total work and \(O(\log m)\) parallel runtime—note that the extra \(\log m\) overhead arises from the obliviousness requirement.
Finally, we apply the loose compaction algorithm in an offline/online fashion too. In the offline phase, we execute Pippenger’s protocol obliviously on a PRAM to compute the satisfying assignment—the offline phase can be parallelized over all recursion depths, thus incurring \(O(\log m)\) parallel runtime overall. In the online phase, we carry out the satisfying assignment that has already been recorded in the offline phase to perform the actual routing of the fetched position labels, and this can be accomplished in O(1) online parallel runtime.
3 Definitions
3.1 Parallel RandomAccess Machines
We review the concepts of a parallel randomaccess machine (PRAM) and an oblivious parallel randomaccess machine (OPRAM). Some of the definitions in this section are borrowed verbatim from Boyle et al. [6] or Chan and Shi [10].
Although we give definitions only for the parallel case, we point out that this is without loss of generality, since a sequential RAM can be thought of as a special case PRAM with one CPU.
Parallel RandomAccess Machine (PRAM). A parallel randomaccess machine consists of a set of CPUs and a shared memory denoted by \({\mathsf {mem}} \) indexed by the address space \(\{0, 1, \ldots , N1\}\), where N is a power of 2. In this paper, we refer to each memory word also as a block, which is at least \(\varOmega (\log N)\) bits long.
In a PRAM, each step of the execution can employ multiple CPUs, and henceforth we use \(m_t\) to denote the number of CPUs involved in executing the tth step for \(t \in \mathbb {N}\). In each step, each CPU executes a next instruction circuit denoted \(\varPi \), updates its CPU state; and further, CPUs interact with memory through request instructions \(\mathbf {I}^{(t)} := (I^{(t)}_i: i \in [m_t])\). Specifically, at time step t, CPU i’s instruction is of the form \(I^{(t)}_i := ({\mathsf {read}}, {\mathsf {addr}})\), or \(I^{(t)}_i := ({\mathsf {write}}, {\mathsf {addr}}, {\mathsf {data}})\) where the operation is performed on the memory block with address \({\mathsf {addr}} \) and the block content \({\mathsf {data}} \).
If \(I^{(t)}_i = ({\mathsf {read}}, {\mathsf {addr}})\) then the \(\text {CPU} \) i should receive the contents of \({\mathsf {mem}} [{\mathsf {addr}} ]\) at the beginning of time step t. Else if \(I^{(t)}_i = ({\mathsf {write}}, {\mathsf {addr}}, {\mathsf {data}})\), \(\text {CPU} \) i should still receive the contents of \({\mathsf {mem}} [{\mathsf {addr}} ]\) at the beginning of time step t; further, at the end of step t, the contents of \({\mathsf {mem}} [{\mathsf {addr}} ]\) should be updated to \({\mathsf {data}} \).
Write Conflict Resolution. By definition, multiple \(\mathsf {read}\) operations can be executed concurrently with other operations even if they visit the same address. However, if multiple concurrent \(\mathsf {write}\) operations visit the same address, a conflict resolution rule will be necessary for our PRAM to be welldefined. In this paper, we assume the following:

The original PRAM supports concurrent reads and concurrent writes (CRCW) with an arbitrary, parametrizable rule for write conflict resolution.

Our compiled, oblivious PRAM (defined below) is a “concurrent read, exclusive write” PRAM (CREW). In other words, our OPRAM algorithm must ensure that there are no concurrent writes at any time.
CPUtoCPU Communication. In the remainder of the paper, we sometimes describe our algorithms using CPUtoCPU communication. For our OPRAM algorithm to be oblivious, the interCPU communication pattern must be oblivious too. We stress that such interCPU communication can be emulated using shared memory reads and writes. Therefore, when we express our performance metrics, we assume that all interCPU communication is implemented with shared memory reads and writes. In this sense, our performance metrics already account for any interCPU communication, and there is no need to have separate metrics that characterize interCPU communication. In contrast, some earlier works [11] adopt separate metrics for interCPU communication.
Additional Assumptions and Notations. Henceforth, we assume that each CPU can only store O(1) memory blocks. Further, we assume for simplicity that the runtime T of the PRAM is fixed a priori and publicly known. Therefore, we can consider a PRAM to be parametrized by the following tuple
where \(\varPi \) denotes the next instruction circuit, N denotes the total memory size (in terms of number of blocks), T denotes the PRAM’s total runtime, and \(m_t\) denotes the number of CPUs in the tth step for \(t \in [T]\).
Finally, in this paper, we consider PRAMs that are stateful and can evaluate a sequence of inputs, carrying state in between. Without loss of generality, we assume each input can be stored in a single memory block.
3.2 Oblivious Parallel RandomAccess Machines
An OPRAM is a (randomized) PRAM with certain security properties, i.e., its access patterns leak no information about the inputs to the PRAM.
Randomized PRAM. A randomized PRAM is a PRAM where the CPUs are allowed to generate private random numbers. For simplicity, we assume that a randomized PRAM has a priori known, deterministic runtime, and that the CPU activation pattern in each time step is also fixed a priori and publicly known.
Memory Access Patterns. Given a PRAM program denoted \(\mathsf{PRAM}\) and a sequence of inputs, we define the notation as follows:

Let T be the total number of parallel steps that \(\mathsf{PRAM}\) takes to evaluate inputs .

Let \(A_t := ({\mathsf {addr}} ^t_1, {\mathsf {addr}} ^t_2, \ldots , {\mathsf {addr}} ^t_{m_t})\) be the list of addresses such that the ith CPU accesses memory address \({\mathsf {addr}} ^t_i\) in time step t.

We define to be the random object \(\left[ A_t \right] _{t \in [T]}\).
Oblivious PRAM (OPRAM). We say that a \(\mathsf{PRAM}\) is perfectly oblivious, iff for any two input sequences \(\mathsf{inp}_0\) and \(\mathsf{inp}_1\) of equal length, it holds that the following distributions are identically distributed (where \(\equiv \) denotes identically distributed):
We remark that for statistical and computational security, some earlier works [8, 10] presented an adaptive, composable security notion. The perfectly oblivious counterpart of their adaptive, composable notion is equivalent to our notion defined above. In particular, our notion implies security against an adaptive adversary who might choose the input sequence \(\mathsf{inp}\) adaptively over time after having observed partial access patterns of \(\mathsf{PRAM}\).
We say that \(\mathsf{OPRAM}\) is a perfectly oblivious simulation of \(\mathsf{PRAM}\) iff \(\mathsf{OPRAM}\) is perfectly oblivious, and moreover \(\mathsf{OPRAM}(\mathsf{inp})\) is identically distributed as \(\mathsf{PRAM}(\mathsf{inp})\) for any input \(\mathsf{inp}\). In the remainder of the paper, we always assume that the original PRAM has a fixed number of CPUs (denoted m) in all steps of execution. For the compiled OPRAM, we consider two models (1) when the OPRAM always consumes exactly m CPUs in every step (i.e., the same number of CPUs as the original PRAM); and (2) when the OPRAM can consume an unbounded number of CPUs in every step; in this case, the actual number of CPUs consumed in each step may vary. We leave it as an open problem how to obliviously simulate a PRAM with a varying number of CPUs (without naïvely padding the number of CPUs to the maximum which can incur large overhead).
Oblivious Simulation Metrics. We adopt the following metrics to characterize the overhead of (parallel) oblivious simulation of a PRAM. In the following, when we say that an OPRAM scheme consumes T parallel steps (or W total work), we mean that the OPRAM scheme consumes T parallel steps (or W total work) except with negligible in N probability. In other words, the definition of our metrics allows the OPRAM to sometimes, but with negligibly small (in N) probability, exceed the desired runtime or total work bound; however, note that the security or correctness failure probability must be 0^{Footnote 6}.

\(\underline{Simulation\,\, overhead}\) (when the OPRAM consumes the same number of CPUs as the PRAM). If a \(\mathsf{PRAM}\) that consumes m CPUs and completes in T parallel steps can be obliviously simulated by an \(\mathsf{OPRAM}\) that completes in \(\gamma \cdot T\) steps also with m CPUs (i.e., the same number of CPUs as the original \(\mathsf{PRAM}\)), then we say that the simulation overhead is \(\gamma \). Note that this means that every PRAM step is simulated by on average \(\gamma \) OPRAM steps.

\(\underline{Total \,\,work \,\,blowup}\) (when the OPRAM may consume unbounded number of CPUs). A \(\mathsf{PRAM}\)’s total work is the number of steps necessary to simulate the \(\mathsf{PRAM}\) under a single CPU, and is equal to the sum \(\sum _{t \in [T]} m_t\). If a \(\mathsf{PRAM}\) of total work W can be obliviously simulated by an \(\mathsf{OPRAM}\) of total work \(\gamma \cdot W\) we say that the total work blowup of the oblivious simulation is \(\gamma \).

\(\underline{Depth\,\, blowup}\) (when the OPRAM may consume unbounded number of CPUs). A \(\mathsf{PRAM}\)’s depth is defined to be its parallel runtime when there are an unbounded number of CPUs. If a \(\mathsf{PRAM}\) of depth D can be obliviously simulated by an \(\mathsf{OPRAM}\) of depth \(\gamma \cdot D\) we say that the depth blowup of the oblivious simulation is \(\gamma \).
Note that the simulation overhead is a good standalone metric if we assume that the OPRAM must consume the same number of CPUs as the PRAM. If the OPRAM is allowed to consume more CPUs than the PRAM, we typically use the metrics total work blowup and depth blowup in conjunction with each other: total work blowup alone does not characterize how much the OPRAM preserves parallelism; and depth blowup alone does not capture the extent to which the OPRAM preserves total work.
Finally, the following simple fact is useful for understanding the complexity of (oblivious) parallel algorithms.
Fact 2
Let \(C > 1\). If an (oblivious) parallel algorithm can complete in T steps consuming m CPUs, then it can complete in CT steps consuming \(\lceil \frac{m}{C} \rceil \) CPUs.
3.3 Building Blocks
In our constructions, we use several useful building blocks such as oblivious routing, oblivious select, oblivious random permutation, etc. Due to lack of space, we describe these building blocks in detail in the full version of the paper [9].
4 Parallel OneTime Oblivious Memory
We define and construct an abstract datatype to process nonrecurrent memory lookup requests. Although the abstraction is similar to the oblivious hashing scheme in Chan et al. [8], our onetime memory scheme needs to be perfectly secure and does not use a hashing scheme. Furthermore, we assume that every real lookup request is tagged with a correct position label that indicates where the requested block is—in this section, we simply assume that the correct position labels are simply provided during lookup; but later in our full OPRAM scheme, we will use a recursive ORAM/OPRAM technique reminiscent of those used in binarytreebased ORAM/OPRAM schemes [10, 12, 36, 38, 39] such that we can obtain the position label of a block first before fetching the block.
4.1 Definition: OneTime Oblivious Memory
Intuition. We describe the intuition using the sequential special case but our formal presentation later will directly describe the parallel version. An oblivious onetime memory supports three operations: (1) Build, (2) Lookup, and (3) Getall. Build is called once upfront to create the data structure: it takes in a set of real blocks (each tagged with its logical address) and creates a data structure that facilitates lookup. After this data structure is created, a sequence of lookup operations can be performed: each lookup can request a real block identified by its logical address or a dummy block denoted \(\bot \) — if the requested block is a real block, we assume that the correct position label is supplied to indicate where in the data structure the requested block is. Finally, when the data structure is no longer needed, one may call a Getall operation to obtain a list of blocks (tagged with their logical addresses) that have not been looked up yet—in our OPRAM scheme later, this is the set of blocks that need to be preserved during rebuilding.
We require that our oblivious onetime memory data structure retain obliviousness as long as (1) the sequence of real blocks looked up all exist in the data structure (i.e., it appeared as part of the input to Build), and moreover, each logical address is looked up at most once; and (2) at most \(\widetilde{n}\) number of dummy lookups may be made where \(\widetilde{n}\) is a predetermined parameter (that the scheme is parametrized with).
Formal Definition. Our formal presentation will directly describe the parallel case. In the parallel version, lookup requests come in batches of size \(m > 1\).
A (parallel) onetime memory scheme denoted \(\mathsf{OTM}^{[n, m, t]}\) is parametrized by three parameters: n denotes the upper bound on the number of real elements; m is the batch size for lookups; t is the upper bound on the number of batch lookups supported. We use three parameters because we use different versions of \(\mathsf{OTM}\). For the basic version in Sect. 5, we have \(t = \frac{n}{m}\) number of batch lookups, whereas for the lowdepth version, the number of batch lookups is larger (which means that some of the lookup addresses must be dummy).
The (parallel) onetime memory scheme \(\mathsf{OTM}^{[n, m, t]}\) is comprised of the following possibly randomized, stateful algorithms to be executed on a ConcurrentRead, ExclusiveWrite PRAM — note that since the algorithms are stateful, every invocation will update an implicit data structure in memory. Henceforth we use the terminology key and value in the formal description but in our OPRAM scheme later, a real key will be a logical memory address and its value is the block’s content.

\(U \leftarrow \mathsf{Build}(\{(k_i, v_i) : i \in [n]\})\): given a set of n keyvalue pairs \((k_i, v_i)\), where each pair is either real or of the form \((\bot , \bot )\), the \(\mathsf{Build}\) algorithm creates an implicit data structure to facilitate subsequent lookup requests, and moreover outputs a list U of exactly n keyposition pairs where each pair is of the form \((k, {\mathsf {pos}})\). Further, every real key input to \(\mathsf{Build}\) will appear exactly once in the list U; and the list U is padded with \(\bot \) to a length n. Note that U does not include the values \(v_i\)’s. Later in our scheme, this keyposition list U will be propagated back to the parent recursion depth during a coordinated rebuild^{Footnote 7}.

\((v_i : i \in [m]) \leftarrow \mathsf{Lookup}(\{(k_i, {\mathsf {pos}} _i): i \in [m]\})\): there are m concurrent \(\mathsf{Lookup}\) operations in a single batch, where we allow each key \(k_i\) requested to be either real or \(\bot \). Moreover, in each batch, at most n / t of the keys are real.

\(R \leftarrow \mathsf{Getall}\): the \(\mathsf{Getall}\) algorithm returns an array R of length n where each entry is either \(\bot \) or real and of the form (k, v). The array R should contain all real entries that have been inserted during \(\mathsf {Build}\) but have not been looked up yet, padded with \(\bot \) to a length of n.
Valid Request Sequence. Our oblivious onetime memory ensures obliviousness only if lookups are nonrecurrent (i.e., never look for the same real key twice); and moreover the number of lookups requests must be upper bounded by a predetermined parameter. More formally, a sequence of operations is valid, iff the following holds:

The sequence begins with a single call to \(\mathsf{Build}\) upfront; followed by a sequence of at most t batch \(\mathsf{Lookup}\) calls, each of which supplies a batch of m keys and the corresponding position labels; and finally the sequence ends with a single call to \(\mathsf{Getall}\).

The \(\mathsf{Build}\) call is supplied with an input array \(S := \{(k_i, v_i)\}_{i \in [n]}\), such that any two real entries in S must have distinct keys.

For every \(\mathsf{Lookup}(\{(k_i, {\mathsf {pos}} _i) : i \in [m]\})\) query in the sequence, if each \(k_i\) is a real key, then \(k_i\) must be contained in S that was input to \(\mathsf{Build}\) earlier. In other words, \(\mathsf{Lookup}\) requests are not supposed to ask for real keys that do not exist in the data structure^{Footnote 8}; moreover, each \((k_i, {\mathsf {pos}} _i)\) pair supplied to \(\mathsf{Lookup}\) must exist in the U array returned by the earlier invocation of \(\mathsf{Build}\), i.e., \({\mathsf {pos}} _i\) must be a correct position label for \(k_i\); and

Finally, in all \(\mathsf{Lookup}\) requests in the sequence, no two keys requested (either in the same or different batches) are the same.
Correctness. Correctness requires that

1.
for any valid request sequence, with probability 1, for every \(\mathsf{Lookup}(\{(k_i, {\mathsf {pos}} _i): i \in [m]\})\) request, the ith answer returned must be \(\bot \) if \(k_i = \bot \); else if \(k_i \ne \bot \), \(\mathsf{Lookup}\) must return the correct value \(v_i\) associated with \(k_i\) that was input to the earlier invocation of \(\mathsf{Build}\).

2.
for any valid request sequence, with probability 1, \(\mathsf{Getall}\) must return an array R containing every (k, v) pair that was supplied to \(\mathsf{Build}\) but has not been looked up; moreover the remaining entries in R must all be \(\bot \).
Perfect Obliviousness. We say that two valid request sequences are lengthequivalent, if the input sets to \(\mathsf{Build}\) have equal size, and the number of \(\mathsf{Lookup}\) requests (where each request asks for a batch of m keys) in the two sequences are equal.
We say that a (parallel) onetime memory scheme is perfectly oblivious, iff for any two lengthequivalent request sequences that are valid, the distribution of access patterns resulting from the algorithms are identically distributed.
4.2 Construction
Intuition. We first explain the intuition for the sequential case, i.e., \(m = 1\). The intuition is simply to permute all elements received as input during \(\mathsf{Build}\). However, since subsequent lookup requests may be dummy (also denoted \(\bot \)), we also need to pad the array with sufficiently many dummies to support these lookup requests. The important invariant is that each real element as well as each dummy will be accessed at most once during lookup requests. For reals, this is guaranteed since the definition of a valid request sequence requires that each real key be requested no more than once, and that each real key requested must exist in the data structure. For dummies, every time a \(\bot \)request is received, we always look for an unvisited dummy. To implement this idea, one tricky detail is that unlike real lookup requests, dummy requests do not carry the position label of the next dummy to be read—thus our data structure itself must maintain an oblivious linked list of dummies such that we can easily find out where the next dummy is. Since all real and dummies are randomly permuted during \(\mathsf{Build}\), and due to the aforementioned invariant, every lookup visits a completely random location of the data structure thus maintaining perfect obliviousness.
It is not too difficult to make the above algorithm parallel (i.e., for the case \(m > 1\)). To achieve this, one necessary modification is that instead of maintaining a single dummy linked list, we now must maintain m dummy linked lists. These m dummy linked lists are created during \(\mathsf{Build}\) and consumed during \(\mathsf{Lookup}\).
Detailed Construction. At the end of Build, our algorithm creates an inmemory data structure consisting of the following:

1.
An array A of length \(n + \widetilde{n}\), where \(\widetilde{n} := tm\) denotes the number of dummies and n denotes the number of real elements. Each entry of the array A (real or dummy alike) has four fields \(({\mathsf {key}}, {\mathsf {val}}, {\mathsf {next}}, {\mathsf {pos}})\) where

\({\mathsf {key}} \) is a key that is either real or dummy; and \({\mathsf {val}} \) is a value that is either real or dummy.

the field \({\mathsf {next}} \in [0..n + \widetilde{n})\) matters only for dummy entries, and at the end of the Build algorithm, the \({\mathsf {next}} \) field stores the position of the next entry in the dummy linked list (recall that all dummy entries form m linked lists); and

the field \({\mathsf {pos}} \in [0..n + \widetilde{n})\) denotes where in the array an entry finally wants to be—at the end of the Build algorithm it must be that \(A[i].{\mathsf {pos}} = i\). However, during the algorithm, entries of A will be permuted transiently; but as soon as each element i has decided where it wants to be (i.e., \(A[i].{\mathsf {pos}} \)), it will always carry its desired position around during the remainder of the algorithm.


2.
An array that stores the head pointers of all m dummy linked lists. Specifically, we denote the m head pointers as \(\{{\mathsf {dpos}} _i: i \in [m]\}\) where each \({\mathsf {dpos}} _i \in [0..n + \widetilde{n})\) is the head pointer of one dummy linked list.
These inmemory data structures, including A and the dummy pointers will then be updated during \(\mathsf{Lookup}\).
\(\underline{Build.}\) Our oblivious \(\mathsf{Build}(\{(k_i, v_i)\}_{i \in [n]})\) algorithm proceeds as follows.

1.
Initialize. Construct an array A of length \(n + \widetilde{n}\) whose entries are of the form described above. Specifically, the keys and values for the first n entries of A are copied from the input. Recall that the input may contain dummies too, and we use \(\bot \) to denote a dummy key from the input.
The last \(\widetilde{n}\) entries of A contain special dummy keys that are numbered. Specifically, for each \(i \in [1..\widetilde{n}]\), we denote \(A_n[i] := A[n1 + i]\), and the entry stored at \(A_n[i]\) has key \(\bot _i\) and value \(\bot \).

2.
Every element decides at random its desired final position. Specifically, perform a perfectly oblivious random permutation on the entries of A—this random permutation decides where each element finally wants to be.
Now, for each \(i \in [0..n + \widetilde{n})\), let \(A[i].{\mathsf {pos}}:= i\). At this moment, \(A[i].{\mathsf {pos}} \) denotes where the element A[i] finally wants to be. Henceforth in the algorithm, the entries of A will be moved around but each element always carries around its desired final position.

3.
Construct the keyposition map U. Perform oblivious sorting on A using the field \({\mathsf {key}} \). We assume that real keys have the highest priority followed by \(\bot< \bot _1< \cdots < \bot _{\widetilde{n}}\) (where smaller keys come earlier).
At this moment, we can construct the keyposition map U from the first n entries of A—recall that each entry of U is of the form \((k, {\mathsf {pos}})\).

4.
Construct m dummy linked lists. Observe that the last \(\widetilde{n}\) entries of A contain special dummy keys, on which we perform the following to build m disjoint singlylinked lists (each of which has length t). For each \(i \in [1..\widetilde{n}]\), if \(i \mod t \ne 0\) we update the entry \(A_n[i].{\mathsf {next}}:= A_n[i+1].{\mathsf {pos}} \), i.e., each dummy entry (except the last entry of each linked list) records its next pointer.
We next record the positions of the heads of the m lists. For each \(i \in [m]\), we set \({\mathsf {dpos}} _i := A_n[t(i1)].{\mathsf {pos}} \).

5.
Move entries to their desired positions. Perform an oblivious sort on A, using the fourth field \({\mathsf {pos}} \). (This restores the ordering according to the previous random permutation.)
At this moment, the data structure \((A, \{{\mathsf {dpos}} _i: i \in [m]\})\) is stored in memory. The keyposition map U is explicitly output and later in our OPRAM scheme it will be passed to the parent recursion depth during coordinated rebuild.
Fact 3
Consuming \(O(\widetilde{n} + n)\) CPUs and setting \((\widetilde{n} + n)^2 \le \lambda \le 2^{\widetilde{n} + n}\), the \(\mathsf{Build}\) algorithm completes in \(O(\log (\widetilde{n} + n))\) parallel steps, except with probability negligible in \(\lambda \).
\(\underline{Lookup.}\) We implement a batch of m concurrent lookup operations \(\{\mathsf{Lookup}(\{(k_i, {\mathsf {pos}} _i): i \in [m]\})\) as follows. For each \(i \in [m]\), we perform the following in parallel.

1.
Decide position to fetch from. If \(k_i \ne \bot \) is real, set \({\mathsf {pos}}:= {\mathsf {pos}} _i\), i.e., we want to use the position label supplied from the input. Else if \(k_i = \bot \), set \({\mathsf {pos}}:= {\mathsf {dpos}} _i\), i.e., the position to fetch from is the next dummy in the ith dummy linked lists. (To ensure obliviousness, the algorithm can always pretend to execute both branches of the ifstatement.)
At this moment, \({\mathsf {pos}} \) is the position to fetch from (for the ith request out of m concurrent requests).

2.
Read and remove. Read the value from \(A[{\mathsf {pos}} ]\) and mark \(A[{\mathsf {pos}} ] := \bot \).

3.
Update dummy head pointer if necessary. If \({\mathsf {pos}} = {\mathsf {dpos}} _i\), update the dummy head pointer \({\mathsf {dpos}} _i := {\mathsf {next}} \). (To ensure obliviousness, the algorithm can pretend to modify \({\mathsf {dpos}} _i\) in any case.)

4.
Return. Return the value read in the above Step 4.2.
The following fact is straightforward from the description of the algorithm.
Fact 4
The \(\mathsf{Lookup}\) algorithm completes in O(1) parallel steps with O(m) CPUs.
\(\underline{Getall.}\) \(\mathsf{Getall}\) is implemented by the following simple procedure: obliviously sort A by the key such that all real entries are packed in front. Return the first n entries of the resulting array (and removing the metadata entries \({\mathsf {next}} \) and \({\mathsf {pos}} \)).
Fact 5
The \(\mathsf{Getall}\) algorithm completes in \(\log (\widetilde{n} + n)\) parallel steps consuming \(O(\widetilde{n} + n)\) CPUs.
Lemma 1
(Perfect obliviousness of the onetime memory scheme). The above (parallel) onetime memory scheme satisfies perfect obliviousness.
Due to lack of space, we defer the proof to the full version of the paper [9].
Summarizing the above, we conclude with the following theorem.
Theorem 6
(Onetime oblivious memory). Let \(\lambda \in \mathbb {N}\) be a parameter related to the probability that the algorithm’s runtime exceeds a desired bound. Assume that each memory block can store at least \(\log n + \log \lambda \) bits. There exists a perfectly oblivious onetime scheme such that \(\mathsf{Build}\) takes \(O(\log n)\) parallel steps (except with negligible in \(\lambda \) probability) consuming n CPUs, \(\mathsf{Lookup}\) for a batch of m requests takes O(1) parallel steps consuming m CPUs, and \(\mathsf{Getall}\) takes \(O(\log n)\) parallel steps consuming n CPUs.
5 Basic OPRAM with \(O(\log ^3 N)\) Simulation Overhead
Recall that N denotes the number of logical memory blocks consumed by the original PRAM, and each memory block can store at least \(\varOmega (\log N)\) bits. In this section, we describe an OPRAM construction such that each batch of m memory requests takes \(O(\log ^3 N)\) parallel steps to satisfy with m CPUs. In the full version of our paper [9], we will describe how to further parallelize the OPRAM when the OPRAM can consume more CPUs than the original PRAM.
Roadmap. We briefly explain the technical roadmap of this section:

In Sect. 5.1, we will first describe a positionbased OPRAM that supports two operations: Lookup and Shuffle. A positionbased OPRAM is an almost fully functional OPRAM scheme except that every real lookup request must supply a correct position label. In our OPRAM construction, these position labels will have been fetched from small recursion depths and therefore will be ready when looking up the positionbased OPRAM.
Our positionbased OPRAM relies on the hierarcial structure proposed by Goldreich and Ostrovsky [21, 22], as well as techniques by Chan et al. [8] that showed how to parallelize such a hierarchical framework.

In Sect. 5.2, we explain how to leverage “coordinated rebuild” and recursion techniques to build a recursive OPRAM scheme that composes logarithmically many instances of our positionbased OPRAM, of geometrically decreasing sizes.
5.1 PositionBased OPRAM
Our basic OPRAM scheme (Sect. 5.2) will consist of logarithmically many positionbased OPRAMs of geometrically increasing sizes, henceforth denoted \(\mathsf{OPRAM}_0\), \(\mathsf{OPRAM}_1\), \(\mathsf{OPRAM}_2\), \(\ldots \), \(\mathsf{OPRAM}_{D}\) where \(D := \log _2 N  \log _2 m\). Specifically, \(\mathsf{OPRAM}_d\) stores \(\varTheta (2^d \cdot m)\) blocks where \(d \in \{0, 1, \ldots , D\}\). The last one \(\mathsf{OPRAM}_D\) stores the actual data blocks whereas every other \(\mathsf{OPRAM}_d\) where \(d < D\) recursively stores the position labels for the next depth \(d+1\).
Data Structure. As we shall see, the case \(\mathsf{OPRAM}_0\) is trivial and is treated specially at the end of this section (Sect. 5.1). Below we focus on describing \(\mathsf{OPRAM}_d\) for some \(1 \le d \le D = \log N  \log m\). For \(d \ne 0\), each \(\mathsf{OPRAM}_d\) consists of \(d+1\) levels geometrically growing in size, where each level is a onetime oblivious memory scheme as defined and described in Sect. 4. We specify this data structure more formally below.
Hierarchical Levels. The positionbased \(\mathsf{OPRAM}_d\) consists of \(d + 1\) levels henceforth denoted as \((\mathsf{OTM}_j: j = 0, \ldots , d)\) where level j is a onetime oblivious memory scheme,
with at most \(n = 2^j \cdot m\) real blocks and m concurrent lookups in each batch (which can all be real). This means that for every \(\mathsf{OPRAM}_d\), the smallest level is capable of storing up to m real blocks. Every subsequent level can store twice as many real blocks as the previous level. For the largest \(\mathsf{OPRAM}_D\), its largest level is capable of storing N real blocks given that \(D = \log N  \log m\)—this means that the total space consumed is O(N).
Every level j is marked as either empty (when the corresponding \(\mathsf{OTM}_j\) has not been rebuilt) or full (when \(\mathsf{OTM}_j\) is ready and in operation). Initially, all levels are marked as empty, i.e., the OPRAM initially is empty.
Position Label. Henceforth we assume that a position label of a block specifies (1) which level the block resides in; and (2) the position within the level the block resides at.
Additional Assumption. We assume that each block is of the form (logical address, payload), i.e., each block carries its own logical address.
Operations. Each positionbased OPRAM supports two operations, \(\mathsf{Lookup}\) and \(\mathsf{Shuffle}\). For every \(\mathsf{OPRAM}_d\) consisting of \(d+1\) levels, we rely on the following algorithms for \(\mathsf{Lookup}\) and \(\mathsf{Shuffle}\).
\(\underline{Lookup.}\) Every batch lookup operation, denoted \(\mathsf{Lookup}(\{({\mathsf {addr}} _i, {\mathsf {pos}} _i) : i \in [m]\})\) receives as input the logical addresses of m blocks as well as a correct position label for each requested block. To complete the batch lookup request, we perform the following.

1.
For each level \(j = 0, \ldots , d\) in parallel, perform the following:

For each \(i \in [m]\) in parallel, first check the supplied position label \({\mathsf {pos}} _i\) to see if the requested block resides in the current level j: if so, let \({\mathsf {addr}} '_i := {\mathsf {addr}} _i\) and let \({\mathsf {pos}} '_i := {\mathsf {pos}} _i\) (and specifically the part of the position label denoting the offset within level j); else, set \({\mathsf {addr}} '_i := \bot \) and \({\mathsf {pos}} '_i := \bot \) to indicate that this should be a dummy request.

\((v_{ij}: i \in [m]) \leftarrow \mathsf{OTM}_j.\mathsf{Lookup}(\{{\mathsf {addr}} '_i, {\mathsf {pos}} '_i : i \in [m]\})\).


2.
At this point, each of the m CPUs has d answers from the d levels respectively, and only one of them is the valid answer. Now each of the m CPUs chooses the correct answer as follows.
For each \(i \in [m]\) in parallel: set \({\mathsf {val}} _i\) to be the only nondummy element in \((v_{ij}: j = 0, \ldots , d)\), if it exists; otherwise set \({\mathsf {val}} _i := \bot \). This step can be accomplished using an oblivious select operation in \(\log d\) parallel steps consuming d CPUs.

3.
Return \(({\mathsf {val}} _i : i \in [m])\).
We remark that in Goldreich and Ostrovsky’s original hierarchical ORAM [21, 22], the hierarchical levels must be visited sequentially—for obliviousness, if the block is found in some smaller level, all subsequent levels must perform a dummy lookup. Here we can visit all levels in parallel since the position label already tells us which level it is in. Now the following fact is straightforward:
Fact 7
For \(\mathsf{OPRAM}_d\), \(\mathsf{Lookup}\) consumes \(O(\log d)\) parallel steps consuming \(m \cdot d\) CPUs where m is the batch size.
\(\underline{Shuffle.}\) Similar to earlier hierarchical ORAMs [21, 22] and OPRAMs [8], a shuffle operation merges consecutively full levels into the next empty level (or the largest level). However, in our Shuffle abstraction, there is an input U that contains some logical addresses together with new values to be updated. Moreover, the shuffle operation is associated with an update function that determines how the new values in U should be incorporated into the OTM during the rebuild.
In our full OPRAM scheme later, the update array U will be passed from the immediate next depth \(\mathsf{OPRAM}_{d+1}\), and contains the new position labels that \(\mathsf{OPRAM}_{d+1}\) has chosen for recently accessed logical addresses. These position labels must then be recorded by \(\mathsf{OPRAM}_d\) appropriately.
More formally, each positionbased \(\mathsf{OPRAM}_d\) supports a shuffle operation, denoted \(\mathsf{Shuffle}(U, \ell ; \mathsf{update})\), where the parameters are explained as follows:

1.
An update array U in which each (nondummy) entry contains a logical address that needs to be updated, and a new value for this block. (Strictly speaking, we allow a block to be partially updated.)
We will define additional constraints on U subsequently.

2.
The level \(\ell \) to be rebuilt during this shuffle.

3.
An \(\mathsf{update}\) function that specifies how the information in U is used to compute the new value of a block in the OTM.
The reason we make this rule explicit in the notation is that a block whose address that appears in U may only be partially modified; hence, we later need to specify this update function carefully. However, to avoid cumbersome notation, we may omit the parameter \(\mathsf{update}\), and just write \(\mathsf{Shuffle}(U,\ell )\), when the context is clear.
For each \(\mathsf{OPRAM}_d\), when \(\mathsf{Shuffle}(U, \ell ; \mathsf{update})\) is called, it must be guaranteed that \(\ell \le d\); and moreover, either level \(\ell \) must either be empty or \(\ell = d\) (i.e., this is the largest level in \(\mathsf{OPRAM}_d\)). Moreover, there is an extra \(\mathsf{OTM}'_0\); jumping ahead, we shall see that \(\mathsf{OTM}'_0\) contains the blocks that are freshly fetched.
The \(\mathsf{Shuffle}\) algorithm then combines levels \(0, 1, \ldots , \ell \) (of \(\mathsf{OPRAM}_d\)), together with the extra \(\mathsf{OTM}'_0\), into level \(\ell \), updating some blocks’ contents as instructed by the update array U and the update function \(\mathsf{update}\). At the end of the shuffle operation, all levels \(0, 1, \ldots , \ell 1\) are now marked as empty and level \(\ell \) is now marked as full.
We now explain the assumptions we make on the update array U and how we want the update procedure to happen:

We require that each logical address appears at most once in U.

Let A be all logical addresses remaining in levels 0 to \(\ell \) in \(\mathsf{OPRAM}_d\): it must hold that the set of logical addresses in U is a subset of those in A. In other words, a subset of the logical addresses in A will be updated before rebuilding level \(\ell \).

If some logical address \({\mathsf {addr}} \) exists only in A but not in U, after rebuilding level \(\ell \), the block’s value from the current \(\mathsf{OPRAM}_d\) should be preserved. If some logical address \({\mathsf {addr}} \) exists in both A and in U, we use the update function to modify its value: update takes a pair of blocks \(({\mathsf {addr}}, {\mathsf {data}})\) and \(({\mathsf {addr}}, {\mathsf {data}} ')\) with the same address but possibly different contents (the first of which coming from the current \(\mathsf{OPRAM}_d\) and the second coming from U), and computes the new block content \({\mathsf {data}} ^*\) appropriately.
We remark that the new value \({\mathsf {data}} ^*\) might depend on both \({\mathsf {data}} \) and \({\mathsf {data}} '\). Later, we will describe how the \(\mathsf{update}\) rule is implemented.
Upon receiving \(\mathsf{Shuffle}(U, \ell ; \mathsf{update})\), proceed with the following steps:

1.
Let \(A :=\cup _{i=0}^\ell \mathsf{OTM}_i.\mathsf{Getall} \cup \mathsf{OTM}'_0.\mathsf{Getall} \), where the operator \(\cup \) denotes concatenation. Moreover, for an entry in A that comes from \(\mathsf{OTM}_i\), then it also carries a label i.
At this moment, the old \(\mathsf{OTM}_0, \ldots , \mathsf{OTM}_\ell \) instances may be destroyed.

2.
We obliviously sort \(A \cup U\) in increasing order of logical addresses, and moreover, placing all dummy entries at the end. If two blocks have the same logical address, place the entry coming from A in front of the one coming from U.
At this moment, in one linear scan, we can operate on every adjacent pair of entries using the aforementioned update operation, such that if they share the same logical address, the first entry is preserved and updated to a new value, and the second entry is set to dummy.
We now obliviously sort the resulting array moving all dummies to the end. We truncate the resulting array preserving only the first \(2^\ell \cdot m\) elements and let \(A'\) denote the outcome (note that only dummies and no real blocks will truncated in the above step).

3.
Next, we call \(U' \leftarrow \mathsf{Build}(A')\) that builds a new \(\mathsf{OTM}'\) and \(U'\) contains the positions of blocks in \(\mathsf{OTM}'\).

4.
\(\mathsf{OTM}'\) is now the new level \(\ell \) and henceforth it will be denoted \(\mathsf{OTM}_\ell \). Mark level \(\ell \) as full and levels \(0, 1, \ldots , \ell 1\) as empty. Finally, output \(U'\) (in our full OPRAM construction, \(U'\) will be passed to the immediately smaller positionbased OPRAM as the update array for performing its shuffle).
If we realize the oblivious sort with the AKS network [1] that sorts n items in \(O(\log n)\) parallel steps consuming n CPUs, we easily obtain the following fact—note that there is a negligible in N probability that the algorithm runs longer than the stated asymptotic time due to the oblivious random permutation building block.
Fact 8
Suppose that the \(\mathsf{update}\) function can be evaluated by a single CPU in O(1) steps. For \(\mathsf{OPRAM}_d\), let \(\ell \le d\), then except with negligible in N probability, \(\mathsf{Shuffle}(U, \ell )\) takes \(O(\log (m \cdot 2^\ell ))\) parallel steps consuming \(m \cdot 2^\ell \) CPUs.
Observe that in the above fact, the randomness comes from the oblivious random permutation subroutine used in building the onetime oblivious memory data structure.
Trivial Case: \(\mathsf{OPRAM}_0\). In this case, \(\mathsf{OPRAM}_0\) simply stores its entries in an array A[0..m) of size m and we assume that the entries are indexed by a \((\log _2 m)\)bit string. Moreover, each address is also a (\(\log _2 m\))bit string, whose block is stored at the corresponding entry in A.

Lookup. Upon receiving a batch of m depthm truncated addresses where all the real addresses are distinct, use oblivious routing to route A[0..m) to the requested addresses. This can be accomplished in \(O(m\log m)\) total work and \(O(\log m)\) depth. Note that \(\mathsf{OPRAM}_0\)’s lookup does not receive any position labels.

Shuffle. Since there is only one array A (at level 0), \(\mathsf{Shuffle}(U, 0)\) can be implemented by oblivious sorting.
5.2 OPRAM Scheme from PositionBased OPRAM
Recursive OPRAMs. The OPRAM scheme consists of \(D+1\) positionbased OPRAMs henceforth denoted as \(\mathsf{OPRAM}_0, \mathsf{OPRAM}_1, \mathsf{OPRAM}_2, \ldots , \mathsf{OPRAM}_D\). \(\mathsf{OPRAM}_D\) stores the actual data blocks, whereas every other \(\mathsf{OPRAM}_d\) where \(d \ne D\) recursively stores the position labels for the next data structure \(\mathsf{OPRAM}_{d + 1}\). Our construction is in essence recursive although in presentation we shall spell out the recursion for clarity. Henceforth we often say that \(\mathsf{OPRAM}_d\) is at recursion depth d or simply depth d.
Although we are inspired by the recursion technique for treebased ORAMs [36], using this recursion technique in the context of hierarchical ORAMs/OPRAMs raises new challenges. In particular, we cannot use the recursion in a blackbox fashion like in treebased constructions since all of our (positionbased, hierarchical) OPRAMs must reshuffle in sync with each other in a nonblackbox fashion as will become clear later.
Format of Depthd Block and Address. Suppose that a block’s logical address is a \(\log _2 N\)bit string denoted \({\mathsf {addr}^{\langle D \rangle }} := {\mathsf {addr}} [1..(\log _2 N)]\) (expressed in binary format), where \({\mathsf {addr}} [1]\) is the most significant bit. In general, at depth d, an address \({\mathsf {addr}^{\langle d \rangle }} \) is the length\((\log _2 m + d)\) prefix of the full address \({\mathsf {addr}^{\langle D \rangle }} \). Henceforth, we refer to \({\mathsf {addr}^{\langle d \rangle }} \) as a depthd address (or the depthd truncation of \({\mathsf {addr}} \)).
When we look up a data block, we would look up the full address \({\mathsf {addr}^{\langle D \rangle }} \) in recursion depth D; we look up \({\mathsf {addr}^{\langle D1 \rangle }} \) at depth \(D1\), \({\mathsf {addr}^{\langle D2 \rangle }} \) at depth \(D2\), and so on. Finally at depth 0, the \(\log _2 m\)bit address uniquely determines one of the m blocks stored at \(\mathsf{OPRAM}_0\). Since each batch consists of m concurrent lookups, one of them will be responsible for this block in \(\mathsf{OPRAM}_0\).
A block with the address \({\mathsf {addr}^{\langle d \rangle }} \) in \(\mathsf{OPRAM}_d\) stores the position labels for two blocks in \(\mathsf{OPRAM}_{d+1}\), at addresses \({\mathsf {addr}^{\langle d \rangle }}  0\) and \({\mathsf {addr}^{\langle d \rangle }}  1\) respectively. Henceforth, we say that the two addresses \({\mathsf {addr}^{\langle d \rangle }}  0\) and \({\mathsf {addr}^{\langle d \rangle }}  1\) are siblings to each other; \({\mathsf {addr}^{\langle d \rangle }}  0\) is called the left sibling and \({\mathsf {addr}^{\langle d \rangle }}  1\) is called the right sibling. We say that \({\mathsf {addr}^{\langle d \rangle }}  0\) is the left child of \({\mathsf {addr}^{\langle d \rangle }} \) and \({\mathsf {addr}^{\langle d \rangle }}  1\) is the right child of \({\mathsf {addr}^{\langle d \rangle }} \).
Operations. Each batch contains m requests denoted as \(((\mathsf {op} _i, {\mathsf {addr}} _i, {\mathsf {data}} _i): i \in [m])\), where for \(\mathsf {op} _i = {\mathsf {read}} \), there is no \({\mathsf {data}} _i\). We perform the following steps.

1.
Conflict resolution. For every depth \(d \in \{0, 1, \ldots , D\}\) in parallel, perform oblivious conflict resolution on the depthd truncation of all m addresses requested.
For \(d = D\), we suppress duplicate addresses. If multiple requests collide on addresses, we would prefer a write request over a read request (since write requests also fetch the old memory value back before overwriting it with a new value). In the case of concurrent write operations to the same address, we use the properties of the underlying PRAM to determine which write operation prevails.
For \(0 \le d < D\), after conflict resolution, the m requests for \(\mathsf{OPRAM}_d\) become
$$ (({\mathsf {addr}_{i}^{\langle d \rangle }}, {\mathsf {flags}} _i): i \in [m]), $$where each nondummy depthd truncated address \({\mathsf {addr}_{i}^{\langle d \rangle }} \) is distinct and has a twobit \({\mathsf {flags}} _i\) that indicates whether each of two addresses \(({\mathsf {addr}_{i}^{\langle d \rangle }}  0)\) and \(({\mathsf {addr}_{i}^{\langle d \rangle }} 1)\) is requested in \(\mathsf{OPRAM}_{d+1}\). As noted by earlier works on OPRAM [6, 10, 11], conflict resolution can be completed through O(1) number of oblivious sorting operations. We thus defer the details of the conflict resolution procedure to the full version of the paper [9].

2.
Fetch. For \(d = 0\) to D sequentially, perform the following:

For each \(i \in [m]\) in parallel: let \({\mathsf {addr}_{i}^{\langle d \rangle }} \) be the depthd truncation of \({\mathsf {addr}_{i}^{\langle D \rangle }} \).

Call \(\mathsf{OPRAM}_d.\mathsf{Lookup}\) to look up the depthd addresses \({\mathsf {addr}_{i}^{\langle d \rangle }} \) for all \(i \in [m]\); observe that position labels for the lookups of nondummy addresses will be available from the lookup of the previous \(\mathsf{OPRAM}_{d1}\) for \(d \ge 1\), which is described in the next step. Recall that for \(\mathsf{OPRAM}_0\), no position labels are needed.

If \(d < D\), each lookup from a nondummy \(({\mathsf {addr}_{i}^{\langle d \rangle }}, {\mathsf {flags}} _i)\) will return two positions for the addresses \({\mathsf {addr}_{i}^{\langle d \rangle }} 0\) and \({\mathsf {addr}_{i}^{\langle d \rangle }} 1\) in \(\mathsf{OPRAM}_{d+1}\). The two bits in \({\mathsf {flags}} _i\) will determine whether each of these two position labels are needed in the lookup of \(\mathsf{OPRAM}_{d+1}\).
We can imagine that there are m CPUs at recursion depth \(d+1\) waiting for the position labels corresponding to \(\{{\mathsf {addr}_{i}^{\langle d+1 \rangle }}: i \in [m]\}\). Now, using oblivious routing, the position labels can be delivered to the CPUs at recursion depth \(d + 1\).

If \(d = D\), the outcome of \(\mathsf{Lookup}\) will contain the data blocks fetched.
Recall that conflict resolution was used to suppress duplicate addresses. Hence, oblivious routing can be used to deliver each data block to the corresponding CPUs that request it.

In any case, the freshly fetched blocks are updated if needed in the case of \(d=D\), and are placed in \(\mathsf{OTM}'_0\) in each \(\mathsf{OPRAM}_d\).


3.
Maintain. We first consider depth D. Set depthD’s update array \({U^{\langle D \rangle }} := \emptyset \). Suppose that \(\ell ^{\langle D \rangle }\) is the smallest empty level in \(\mathsf{OPRAM}_D\).
We have the invariant that for all \(0 \le d < D\), if \(\ell ^{\langle D \rangle } < d\), then \(\ell ^{\langle D \rangle }\) is also the smallest empty level in \(\mathsf{OPRAM}_d\).
For \(d := D\) downto 0, do the following:

If \(d < \ell ^{\langle D \rangle }\), set \(\ell := d\); otherwise, set \(\ell := \ell ^{\langle D \rangle }\).

Call \(U \leftarrow \mathsf{OPRAM}_d.\mathsf{Shuffle}({U^{\langle d \rangle }}, \ell ; \mathsf{update})\) where \(\mathsf{update}\) is the following natural function: recall that in \({U^{\langle d \rangle }} \) and \(\mathsf{OPRAM}_{d1}\), each depth\((d1)\) logical address stores the position labels for both children addresses. For each of the child addresses, if \({U^{\langle d \rangle }} \) contains a new position label, choose the new one; otherwise, choose the old label previously in \(\mathsf{OPRAM}_{d1}\).

If \(d \ge 1\), we need to send the updated positions involved in U to depth \(d1\).
We use the \(\mathsf{Convert}\) subroutine to convert U into an update array for depth\((d1)\) addresses, where each entry may pack the position labels for up to two sibling depthd addresses. \(\mathsf{Convert}\) can be realized with O(1) oblivious sorting operations and we defer its detailed presentation to the full version of our paper [9].
Now, set \({U^{\langle d1 \rangle }} \leftarrow \mathsf{Convert}(U, d)\), which will be used in the next iteration for recursion depth \(d1\) to perform its shuffle.

With the above basic OPRAM construction, we can achieve the following theorem whose proof is deferred to the full version of the paper [9].
Theorem 9
The above construction is a perfectly secure OPRAM scheme satisfying the following performance overhead:

When consuming the same number of CPUs as the original PRAM, the scheme incurs \(O(\log ^3 N)\) simulation overhead;

When the OPRAM is allowed to consume an unbounded number of CPUs, the scheme incurs \(O(\log ^3 N)\) total work blowup and \(O((\log m + \log \log N) \log N)\) depth blowup.
In either case, the space blowup is O(1).
Proof
We defer the obliviousness proof and performance analysis to the full version of the paper [9].
Note that at this moment, even for the sequential special case, we already achieve asymptotic savings over Damgård et al. [13] in terms of space consumption. Furthermore, Damgård et al. [13]’s construction is sequential in nature and does not immediately give rise to an OPRAM scheme.
6 Conclusion and Future Work
In this paper, we constructed a perfectly secure OPRAM scheme with \(O(\log ^3 N)\) total work blowup, \(O(\log N \log \log N)\) depth blowup, and O(1) space blowup. To the best of our knowledge our scheme is the first perfectly secure (nontrivial) OPRAM scheme, and even for the sequential special case we asymptotically improve the space overhead relative to Damgård et al. [13]. Prior to our work, the only known perfectly secure ORAM scheme is that by Damgård et al. [13], where they achieve \(O(\log ^3 N)\) simulation overhead and \(O(\log N)\) space blowup. No (nontrivial) OPRAM scheme was known prior to our work, and in particular the scheme by Damgård et al. [13] does not appear amenable to parallelization. Finally, in comparison with known statistically secure OPRAMs [10, 39], our work removes the dependence (in performance) on the security parameter; thus we in fact asymptotically outperform known statistically secure ORAMs [39] and OPRAMs [10] when (sub)exponentially small failure probabilities are required.
Exciting questions remain open for future research:

Are there any separations between the performance of perfectly secure and statistically secure ORAMs/OPRAMs?

Can we construct perfectly secure ORAMs/OPRAMs whose total work blowup matches the best known statistically secure ORAMs/OPRAMs assuming negligible security failures?

Can we construct perfectly secure ORAM/OPRAM schemes whose concrete performance lends to deployment in realworld systems?
Notes
 1.
 2.
As we point out later, randomly permuting real blocks is in fact not sufficient; we also need to allow dummy lookups by introducing an oblivious dummy linked list.
 3.
A similar observation was adopted by Goodrich et al. [24] in their statistically secure ORAM construction.
 4.
In fact, later in our paper, we omit the sequential version and directly present the parallel version of all algorithms.
 5.
In Damgård et al. [13], the shuffle phase incurs an \(O(\log ^3 N)\) depth which is the same as the overhead for accessing a block. Specifically, a \(\log N\) factor arises due to oblivious sorting, a \(\log N\) factor due to the existence of hierarchies, and another \(\log N\) factor due to the extra \(\log N\) dummies stored for every real element. Though an offline/online technique like ours may be conceivable for their scheme, the existence of the extra \(\log N\) dummies makes it inherently hard to improve the depth by another \(\log N\) factor.
 6.
Similarly, the perfectly secure ORAM by Damgård et al. [13] also allowed a negligible small probability for the algorithm to exceed the desired complexity bound but the security or correctness failure probability must be 0.
 7.
Note that we do not explicitly denote the implicit data structure in the output of Build, since the implicit data structure is needed only internally by the current oblivious onetime memory instance. In comparison, U is explicitly output since U will later on be (externally) needed by the parent recursion depth in our OPRAM construction.
 8.
References
Ajtai, M., Komlós, J., Szemerédi, E.: An O(N log N) sorting network. In: Proceedings of the Fifteenth Annual ACM Symposium on Theory of Computing, STOC 1983, pp. 1–9. ACM, New York (1983)
Ajtai, M.: Oblivious RAMs without cryptogrpahic assumptions. In: Proceedings of the FortySecond ACM Symposium on Theory of Computing, STOC 2010, pp. 181–190. ACM, New York (2010)
Asharov, G., Chan, T.H.H., Nayak, K., Pass, R., Ren, L., Shi, E.: Oblivious computation with data locality. IACR Cryptology ePrint Archive 2017:772 (2017)
BenOr, M., Goldwasser, S., Wigderson, A.: Completeness theorems for noncryptographic faulttolerant distributed computation. In: Proceedings of the Twentieth Annual ACM Symposium on Theory of Computing, STOC 1988, pp. 1–10 (1988)
Boyle, E., Chung, K.M., Pass, R.: Largescale secure computation: multiparty computation for (parallel) RAM programs. In: Gennaro, R., Robshaw, M. (eds.) CRYPTO 2015. LNCS, vol. 9216, pp. 742–762. Springer, Heidelberg (2015). https://doi.org/10.1007/9783662480007_36
Boyle, E., Chung, K.M., Pass, R.: Oblivious parallel RAM and applications. In: Kushilevitz, E., Malkin, T. (eds.) TCC 2016. LNCS, vol. 9563, pp. 175–204. Springer, Heidelberg (2016). https://doi.org/10.1007/9783662490990_7
Chan, T.H.H., Chung, K.M., Shi, E.: On the depth of oblivious parallel RAM. In: Asiacrypt (2017)
Chan, T.H.H., Guo, Y., Lin, W.K., Shi, E.: Oblivious hashing revisited, and applications to asymptotically efficient ORAM and OPRAM. In: Asiacrypt (2017)
Chan, T.H.H., Nayak, K., Shi, E.: Perfectly secure oblivious parallel ram. Cryptology ePrint Archive, Report 2018/364 (2018). https://eprint.iacr.org/2018/364
Chan, TH.H., Shi, E.: Circuit OPRAM: a unifying framework for computationally and statistically secure ORAMs and OPRAMs. In: TCC (2017)
Chen, B., Lin, H., Tessaro, S.: Oblivious parallel RAM: improved efficiency and generic constructions. In: Kushilevitz, E., Malkin, T. (eds.) TCC 2016. LNCS, vol. 9563, pp. 205–234. Springer, Heidelberg (2016). https://doi.org/10.1007/9783662490990_8
Chung, K.M., Liu, Z., Pass, R.: Statisticallysecure ORAM with \(\tilde{O}(\log ^2n)\) overhead. In: Asiacrypt (2014)
Damgård, I., Meldgaard, S., Nielsen, J.B.: Perfectly secure oblivious RAM without random oracles. In: Theory of Cryptography Conference (TCC), pp. 144–163 (2011)
Dautrich, J., Stefanov, E., Shi, E.: Burst ORAM: minimizing ORAM response times for bursty access patterns. In: 23rd USENIX Security Symposium (USENIX Security 14), pp. 749–764, San Diego, CA, August 2014. USENIX Association
Demertzis, I., Papadopoulos, D., Papamanthou, C.: Searchable encryption with optimal locality: achieving sublogarithmic read efficiency. Cryptology ePrint Archive, Report 2017/749 (2017). https://eprint.iacr.org/2017/749
Fletcher, C.W., Ren, L., Kwon, A., van Dijk, M., Devadas, S.: Freecursive ORAM: [nearly] free recursion and integrity verification for positionbased oblivious RAM. In: ASPLOS (2015)
Fletcher, C.W., Ren, L., Kwon, A., van Dijk, M., Stefanov, E., Devadas, S.: RAW path ORAM: a lowlatency, lowarea hardware ORAM controller with integrity verification. IACR Cryptol. Eprint Arch. 2014, 431 (2014)
Fletcher, C.W., Ren, L., Yu, X., van Dijk, M., Khan, O., Devadas, S.: Suppressing the oblivious RAM timing channel while making information leakage and program efficiency tradeoffs. In: HPCA, pp. 213–224 (2014)
Genkin, D., Ishai, Y., Weiss, M.: Binary AMD circuits from secure multiparty computation. In: Hirt, M., Smith, A. (eds.) TCC 2016. LNCS, vol. 9985, pp. 336–366. Springer, Heidelberg (2016). https://doi.org/10.1007/9783662536414_14
Gentry, C., Goldman, K.A., Halevi, S., Jutla, C.S., Raykova, M., Wichs, D.: Optimizing ORAM and using it efficiently for secure computation. In: Privacy Enhancing Technologies Symposium (PETS) (2013)
Goldreich, O.: Towards a theory of software protection and simulation by oblivious RAMs. In: ACM Symposium on Theory of Computing (STOC) (1987)
Goldreich, O., Ostrovsky, R.: Software protection and simulation on oblivious RAMs. J ACM 43(3), 431–473 (1996)
Goodrich, M.T., Mitzenmacher, M.: Privacypreserving access of outsourced data via oblivious RAM simulation. In: International Colloquium on Automata, Languages and Programming (ICALP), pp. 576–587 (2011)
Goodrich, M.T., Mitzenmacher, M., Ohrimenko, O., Tamassia, R.: Privacypreserving group data access via stateless oblivious RAM simulation. In: Proceedings of the TwentyThird Annual ACMSIAM Symposium on Discrete Algorithms, SODA 2012, pp. 157–167, Philadelphia, PA, USA (2012). Society for Industrial and Applied Mathematics
Dov Gordon, S., Katz, J., Kolesnikov, V., Krell, F., Malkin, T., Raykova, M., Vahlis, Y.: Secure twoparty computation in sublinear (amortized) time. In: ACM Conference on Computer and Communications Security (CCS) (2012)
Ishai, Y., Kushilevitz, E., Ostrovsky, R., Prabhakaran, M., Sahai, A.: Efficient noninteractive secure computation. In: Paterson, K.G. (ed.) EUROCRYPT 2011. LNCS, vol. 6632, pp. 406–425. Springer, Heidelberg (2011). https://doi.org/10.1007/9783642204654_23
Kushilevitz, E., Lu, S., Ostrovsky, R.: On the (in)security of hashbased oblivious RAM and a new balancing scheme. In: ACMSIAM Symposium on Discrete Algorithms (SODA) (2012)
Liu, C., Harris, A., Maas, M., Hicks, M., Tiwari, M., Shi, E.: Ghostrider: a hardwaresoftware system for memory trace oblivious computation. SIGPLAN Not. 50(4), 87–101 (2015)
Liu, C., Wang, X.S., Nayak, K., Huang, Y., Shi, E.: ObliVM: a programming framework for secure computation. In: 2015 IEEE Symposium on Security and Privacy, SP 2015, San Jose, CA, USA, 17–21 May 2015, pp. 359–376 (2015
Maas, M., Love, E., Stefanov, E., Tiwari, M., Shi, E., Asanovic, K., Kubiatowicz, J., Song, D.: Phantom: practical oblivious computation in a secure processor. In: ACM Conference on Computer and Communications Security (CCS) (2013)
Nayak, K., et al.: HOP: hardware makes obfuscation practical. In: 24th Annual Network and Distributed System Security Symposium, NDSS (2017)
Nayak, K., Katz, J.: An oblivious parallel RAM with \(o(log^2 n)\) parallel runtime blowup. IACR Cryptol. Eprint Arch. 2016, 1141 (2016)
Pippenger, N.: Selfrouting superconcentrators. In: Proceedings of the TwentyFifth Annual ACM Symposium on Theory of Computing, STOC 1993, pp. 355–361. ACM, New York (1993)
Raskin, M., Simkin, M.: Oblivious RAM with small storage overhead. Cryptology ePrint Archive, Report 2018/268 (2018). https://eprint.iacr.org/2018/268
Ren, L., Yu, X., Fletcher, C.W., van Dijk, M., Devadas, S.: Design space exploration and optimization of path oblivious RAM in secure processors. In: ISCA, pp. 571–582 (2013)
Shi, E., Chan, T.H.H., Stefanov, E., Li, M.: Oblivious RAM with \(O(\log ^3 N)\) worstcase cost. In: ASIACRYPT, pp. 197–214 (2011)
Stefanov, E., Shi, E.: Oblivistore: high performance oblivious cloud storage. In: IEEE Symposium on Security and Privacy (S & P) (2013)
Stefanov, E., van Dijk, M., Shi, E., Fletcher, C., Ren, L., Yu, X., Devadas, S.: Path ORAM  an extremely simple oblivious RAM protocol. In: ACM Conference on Computer and Communications Security (CCS) (2013)
Wang, X.S., Chan, TH.H., Shi, E.: Circuit ORAM: on tightness of the GoldreichOstrovsky lower bound. In: ACM CCS (2015)
Williams, P., Sion, R., Tomescu, A.: PrivateFS: a parallel oblivious file system. In: ACM Conference on Computer and Communications Security (CCS) (2012)
Acknowledgments
Kartik Nayak was supported by a Google Ph.D. fellowship. TH. Hubert Chan was supported in part by the Hong Kong RGC under grant 17200418. Elaine Shi was supported in part by NSF award CNS1601879, a Packard Fellowship, and a DARPA Safeware grant (subcontractor under IBM).
We gratefully acknowledge Shai Halevi and Craig Gentry for helpful discussions and for suggesting the use of expander graphs to achieve lowonlinedepth routing of position labels. We are extremely grateful to Bruce Maggs for most patiently explaining Pippenger’s result [33] to us and answering many of our technical questions. We acknowledge KaiMin Chung for many helpful technical discussions regarding perfectly secure ORAM and OPRAM. We thank Ling Ren for many early discussions on perfectly secure ORAMs. We thank Muthuramakrishnan Venkitasubramaniam, Antigoni Polychroniadou, and KaiMin Chung for helpful discussions on the significance of achieving perfect security in cryptographic primitives, and for helpful editorial comments. Elaine Shi is grateful to Bruce Maggs, Bobby Bhattacharjee, KaiMin Chung, and FengHao Liu for their unwavering moral support during the period this research was conducted.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2018 International Association for Cryptologic Research
About this paper
Cite this paper
Chan, TH.H., Nayak, K., Shi, E. (2018). Perfectly Secure Oblivious Parallel RAM. In: Beimel, A., Dziembowski, S. (eds) Theory of Cryptography. TCC 2018. Lecture Notes in Computer Science(), vol 11240. Springer, Cham. https://doi.org/10.1007/9783030038106_23
Download citation
DOI: https://doi.org/10.1007/9783030038106_23
Published:
Publisher Name: Springer, Cham
Print ISBN: 9783030038090
Online ISBN: 9783030038106
eBook Packages: Computer ScienceComputer Science (R0)