A major issue with respect to explaining machine learning algorithms lies in the area of privacy protection: Trust is one of the core problems when dealing with personal, and potentially sensitive, information, especially when the algorithms in place are hard or even impossible to understand. This can be a major risk for acceptance, not only by the end users, like e.g. hospital patients, or generally in safety-critical decision making [10], but also among the expert engineers that are required to train the models, or, in case of an expert-in-the-loop approach, partake in the daily interaction with the expert system [14]. One option is to include risk management practice early in the project to manage such risks [11]. Trust and Privacy are actually a twofold problem in this regard; an example from the medical domain shall illustrate this: The patients need to be able to trust the machine learning environment that their personal data is secured and protected against theft and misuse, but also that the analytical processes working on their data are limited to the selection they have given consent to.
For the expert, on the other hand, there is the need to trust the environment that their input to the system is not manipulated later on. Furthermore, usability is a fundamental factor for successfully integrating experts into AI systems, which, again, requires the designers of the interfaces to understand the fundamentals of the system in place. Here it must be noted that usability and security are often considered fundamental opposites, hence research in the so-called area of usable security [3] is urgently needed.
A topic closely related to the issue of security and privacy, but still different in nature, is the issue of fingerprinting/watermarking information [22]. Many approaches in utilizing data for data driven research face the problem that data must be shared between partners, i.e. data sets are either sent to a central analysis repository for further processing, or directly shared between the partners themselves. While the earlier approach allows for some kind of control over the data by the trusted third party operating the analysis platform, in the later one, the original owner potentially gives up control over the data set. This might not even be a problem with respect to privacy, as the data shared with the other partners will in most cases obey data protection rules as put forth by various regulations, still, this data might be an asset of high (monetary) value. Thus, when sharing the data with other partners, it must be made sure that the data is not further illegally distributed. A typical reactive approach to this problem is the implementation of so-called fingerprints or watermarks; these can also be used to embedded information that helps to detect collusion in deanonymization attacks [13, 21]. Both terms, fingerprinting and watermarking, are often used synonymously by authors, while others differentiate them as watermarks being mechanisms that prove the authenticity and ownership of a data set and fingerprints actually being able to identify the data leak by providing each partner with the same basic set marked with different fingerprints.
Throughout the past decades, watermarking and fingerprinting of information has gained a lot of attention in the research community, most notably regarding the protection of digital rights in the music and movie industries [23]. Approaches for marking data have also been put forth (e.g. [1]) and while a lot of them exist nowadays, most of them only focus on marking whole data sets and fail with partially leaked sets. Thus, in order to provide transparency with respect to privacy, as well as explainability, we propose that a fingerprinting mechanism within data driven research requires the following criteria:
-
1.
Single Record Detection: The detection of the data leak should be possible with only one single leaked (full) record. This is a major obstacle for most algorithms that rely on adding or removing so-called marker-records from the original data set.
-
2.
Collusion Protection: Several partners being issued the same fingerprinted data set might collude in order to extract and remove the fingerprints, or even frame another partner. The fingerprinting algorithm is required to be stable against such kinds of attacks.
-
3.
High Performance: In order to make this protection mechanism usable, it must not require a lot of resources, neither with respect to calculation time (for both, the generation of the fingerprint, as well as the detection), nor with respect to additional storage requirements.
-
4.
Low distortion: The algorithm must not introduce a large amount of additional distortion, thus further reducing the value of the data used in the analysis.
The development of novel techniques in this area is thus another open problem that has a high potential for future research. When developing new solutions contradicting requirements including future improvements in “counter-privacy”, aka. forensics [2], have to be considered.
Last, but not least, the need to understand machine learning algorithms is required to deal with distortion: Due to novel regulations in the European Union, especially the General Data Protection Regulation (GDPR), the protection of privacy has become extremely important and consent for processing personal information has to be asked for rather narrow use cases, i.e. there is no more “general consent”. Thus, research labs tend to consider anonymizing their data, which makes it non-personal information and thus consent-free to use. Still, as it has already been shown [16], many standard anonymization techniques introduce quite a large amount of distortion into the end results of classical machine learning algorithms. In order to overcome this issue, additional research in the area of Privacy Aware Machine Learning (PAML) is needed: The distortion needs to be quantified in order to be able to select the anonymization algorithm/machine learning algorithm pairing that is ideal with respect to the given data set. Explainable AI can be a major enabler for this issue, as understanding decisions would definitely help in understanding and estimating distortions. In addition, algorithms (both, for anonymization and machine learning) need to be adapted in order to reduce the distortion introduced, again, a task where the black-box characteristics of machine learning nowadays is an issue. Thus, explainable AI could be the key to designing solutions that harness the power of machine learning, while guaranteeing privacy at the same time.