Constructing adversarial examples to investigate the plausibility of explanations in deep audio and image classifiers

Given the rise of deep learning and its inherent black-box nature, the desire to interpret these systems and explain their behaviour became increasingly more prominent. The main idea of so-called explainers is to identify which features of particular samples have the most influence on a classifier’s prediction, and present them as explanations. Evaluating explainers, however, is difficult, due to reasons such as a lack of ground truth. In this work, we construct adversarial examples to check the plausibility of explanations, perturbing input deliberately to change a classifier’s prediction. This allows us to investigate whether explainers are able to detect these perturbed regions as the parts of an input that strongly influence a particular classification. Our results from the audio and image domain suggest that the investigated explainers often fail to identify the input regions most relevant for a prediction; hence, it remains questionable whether explanations are useful or potentially misleading.


Introduction
In recent years, a wide variety of explanation methods (''explainers'') have been developed in order to explain the inner workings of deep neural networks, with two large groups being perturbation-and gradient-based explainers. The large number of approaches available makes it hard or even impossible to decide which explanation technique to choose and which explanation to trust. Since there is no agreement on what constitutes a ''good explanation'' [1], an explanation is often simply evaluated based on whether it looks reasonable or matches one's expectation (''confirmation bias'') [2]. This lack of standard evaluation metrics hinders a comparison between different explanation techniques. Even worse, it was shown that different explanation methods disagree [3,4], which suggests that at least some of them do not capture the real inner workings of a system and that explanations produced by an explainer A cannot serve as a ground truth for a newly developed explainer B.
While no standard evaluation procedure for explanations has emerged so far, a few different approaches were proposed in the past. On one hand, these include metrics that quantify the quality of single explanations, e.g. fidelity, consistency, or comprehensibility [5]. These metrics may sometimes be hard to compute (e.g. comprehensibility) and can-in the worst case-give a false sense of trust into a particular explanation (cf. [6]). On the other hand, also approaches for evaluating an explanation algorithm as a whole were proposed, e.g. region perturbation [7], ROAR [8], sanity checks [2], and the CLEVR-XAI benchmark [9]. But again, the assumptions underlying many of these approaches have been questioned in the past (cf. [9,10]).
Previous approaches to compare different explainers either focus on pixel-wise attributions computed by gradient-based explainers (without considering perturbationbased explainers like Local Interpretable Model-agnostic Explanations (LIME)) [2,8,9] or compare LIME to rule-based techniques or other methods, which are often not applicable to domains such as images [11,12]. A comparison between gradient-based and perturbation-based explanations is not straightforward, as the granularity of these explanations is different; as opposed to pixel-wise explanations, LIME explanations consist of groups of pixels, also called super-pixels or segments. Recent work has therefore tried to address evaluation from a different angle, namely by using dataset modifications [13] or adversarial examples [14] as ground truth. As adversarial perturbations lead to changes in a system's original prediction, the reason for such an (erroneous) prediction can be found in these perturbations of the input. Göpfert et al. [14] use so-called localised adversarial attacks, which attack a constrained region of an input image, in order to evaluate classic Saliency [15], Guided Backpropagation [16], and LIME [17], making it one of the first attempts to compare pixel-wise attribution maps with a method such as LIME. The evaluation in Göpfert et al. [14] is based on how well each explainer recovers the affected region as the cause for the adversarial prediction, with LIME being the explainer that performs best in their setting.
In this work, we follow a similar idea and use adversarial perturbations to evaluate the plausibility of explanations in two different domains, namely audio as well as images. We refine and extend the analysis along several dimensions, in order to provide more detailed insights and a fairer comparison for different explainers. In particular, we test explainers using adversarial perturbations spread over either parts or the entirety of a sample. Since we are not restricting perturbations to very dense regions of a sample, we do not favour explanations which are inherently more concentrated (such as LIME).
In a bit more detail, the paper is structured as follows: First, we give an overview over all examined explainers in Sect. 2, as well as the attacks on the audio / image systems in Sect. 3. We then detail our experimental setup in Sect. 4. In Sect. 5, we evaluate LIME on an audio classification task for which we chose a state-of-the-art Singing Voice Detection System (SVDS) [18]. In this set of experiments, we focus on LIME with time-frequency segments (i.e. rectangular subparts of spectrograms) as an explainer, as this is the most prominent method in audio (cf. [19][20][21][22]). Then, in Sect. 6, we evaluate a wider range of explainers, including earlier approaches (e.g. LIME, Saliency and Guided Backpropagation), as previously compared by Göpfert et al., as well as more recent methods (e.g. SmoothGrad [23] or VarGrad [2]) on a variety of pretrained ImageNet classifiers. For the experiments in the image domain, we use more commonly used super-pixels computed by the Simple Linear Iterative Clustering (SLIC) algorithm [24] as segments, as well as rectangular image segments. We also propose to aggregate pixel-wise attributions to such super-pixels or segments. In contrast with Göpfert et al. [14], we propose a non-random baseline based on the magnitude of the adversarial perturbation for all our experiments. We finally discuss our findings and give an outlook in Sect. 7. This article is an expanded version of a conference paper (Praher et al. [6]), with the following new contributions: • We present the first work that compares different explainers for multiple models in the image domain as well as an explainer applied to an audio system, using adversarial perturbations as ground truth; • We propose to aggregate pixel-wise attributions computed by gradient-based explainers for an improved comparison with LIME; • Finally, in contrast with [14], we compare explanations with a stronger, non-random baseline.

Explainers
This section contains a brief description of a variety of methods that are often used to explain the decisions of computational models, all of which we subsequently compare in different experiments. Each of these following approaches is local (explains individual predictions) and post hoc (can be used on a trained model without modifying it). An explainer can be model-specific (it only works for a subset of model classes with specific properties, e.g. gradient-based methods) or model-agnostic (e.g. LIME).
Model-agnostic explainers can be applied to any model type and only require inputs and the corresponding outputs [25]. For more detailed descriptions, we refer the reader to the original papers.

Local interpretable model-agnostic explanations
The main idea of LIME [17] is to approximate the neighbourhood of a prediction f(x), which we want to explain, with a model g that is simpler than f (most commonly, g is linear). To do this, the first step requires the derivation of an interpretable representation depending on the input domain. For images, this can be (perceptually) grouped pixels, also called super-pixels. Let x 2 R d be a sample in the original input domain 1 , and let x 0 2 f0; 1g d 0 be the interpretable representation (0 / 1 denote the absence / presence of interpretable features). We then sample N s instances around x 0 by randomly generating vectors containing 0's or 1's. Each of these generated instances z 0 is subsequently mapped back to z 2 R d , i.e. its representation in the input domain; z can here be imagined to look like the original input x, yet with all parts that equal 0 in z 0 being occluded. All created instances z 0 and their predictions of the original model f(z) are then used to train the explanation model g [17]. A model g most often corresponds to a linear model, e.g.
Note here that samples z are usually additionally weighted based on how close they are to the original input x during training of the model g.
The most typical interpretable representations used in the image domain are super-pixels (i.e. segments) computed with the SLIC [24] algorithm or any standard image segmentation algorithm [17]. One could also use a neural network to predict image segments. In the audio domain, rectangular time-frequency segments were previously proposed [19].
In what follows, we use these rectangular segments when investigating an SVDS; we also adapt this to our experiments in the image domain, using equally sized rectangular image segments. Additionally, we repeat all experiments in the image domain with SLIC segments.

Gradient-based explanations
Next to LIME, we want to look at a different and popular family of explainers, based on gradients. To avoid confusion between the general term saliency maps (often used to refer to the attribution maps that serve as explanations) and the approach ''Saliency Maps'', we are going to use ''Saliency Maps'' to refer to the concrete algorithm and ''attribution maps'' when we talk about the resulting visualisation.
Saliency Maps (Sal) [15] (also called ''Classic Saliency'' [14] or ''Gradient'' [2,8]) are one of the first approaches for visualising the importance of each pixel for a class decision (''which pixels need to be changed the least to affect the class score the most''). This is achieved by taking the derivative of the class score S c with respect to the input image x, i.e.
Guided Backpropagation (GBP) also takes the derivative of the class score S c with respect to the input image, but additionally overrides the backpropagation of the ReLU function in a way to only backpropagate non-negative gradients. This leads to sharper visualisations than previous methods [16]. Integrated Gradients (IG) [26] addresses different weaknesses of previously published methods. It computes the straight-line path from a baseline x 0 (e.g. a black image) to the input of interest x and accumulates all gradients at k different points along that path, i.e.
Note that f here denotes the model we want to interpret.

Smoothing noisy attributions
Attribution maps like the ones in Sect. 2.2 often look ''noisy'' to a human observer [26] as shown in Fig. 1. In the following, we therefore describe different approaches for smoothing attribution maps. Smoothing can be applied to any of the gradient-based explainers, resulting in 9 different combinations (subsequently indicated by a concatenation of their abbreviations, e.g. IG-SG for Integrated Gradients with SmoothGrad). Examples are shown in Appendix in Fig. 9. Following the terminology in [8], we will refer to the smoothed versions of the previously explained standard explainers as ''ensemble explainers''.
SmoothGrad (SG) [23] averages a set of J noisy attribution maps, which are constructed by adding Gaussian noise N ð0; r 2 Þ with standard deviation r to the input of interest x independently J times: SmoothGrad-Squared (SG 2 ) is based on SmoothGrad with the small difference that the each of the J attribution maps is squared before averaging [8], i.e.
VarGrad (VG) is also based on SmoothGrad, but instead of averaging, the variance over J noisy attribution maps is computed [2], i.e.
2.4 Hyper-parameters for the explanations LIME: The hyper-parameters for LIME were chosen based on preliminary experiments. For audio explanations, we occlude segments based on their mean spectrogram value (cf. [22]). We set the number of perturbed samples to N s ¼ 8; 192 for explanations in the audio domain. This value is set to N s ¼ 512 in the image domain as a trade-off between stability [22] and run time. We use an exponential kernel defined on the cosine distance function with kernel width 0.25 for explanations in both domains.
GBP, Sal, IG: Gradient-based explainers have less hyper-parameters that need to be considered. GBP and Sal do not have any hyper-parameters. For IG, we need to select the baseline, for which we choose a black image (all zeroes).
SG, SG 2 , VG: For SG, SG 2 and VG, we set J (the number of noisy attribution maps that are used for smoothing) to 25, as suggested by [8].

Adversarial attacks
To attack different image classifiers, we apply a targeted version of Projected Gradient Descent (PGD), originally proposed by Madry et al. [27]. The targeted PGD attack iteratively changes an input image such that its system loss w.r.t. a new target prediction is decreased, i.e.
Every adversarial examplex is subsequently initialised with the original (clean) image x. In each iteration ep,x is updated based on the sign of the gradient w.r.t. the input (i.e. rx) of the system loss L. Note that f denotes a system, and t denotes a particular target. The perceptibility of an adversarial perturbation is influenced by the step size g, as well as the clipping factor , which is used to ensure that a perturbation stays within the range of ½À; . We use this attack when attacking image classifiers to avoid unnecessary high runtime complexity (cf. exemplary adversaries in Appendix B.2). For audio data, it was previously shown [28] that additional restrictions when using a Carlini & Wagner (C&W) like attack as opposed to PGD leads to less perceptible audio perturbations. We therefore apply C &W to attack the SVDS. This results in a modified adversarial objective (cf. [28,29]), now expressed as an optimisation for the adversarial perturbation d (i.e. x ¼ x þ d) itself: The notation here remains the same; the most significant change is the integration of the squared L2-norm of the perturbation kdk 2 2 as an attempt to keep it as small (imperceptible) as possible. The factor a balances the focus of the attack on either keeping a perturbation as small as possible, or changing the prediction on the adversary quickly.
To find hyper-parameters for PGD, we perform a grid search over values for the clipping factor and the update rate g on a subset of our image data. We choose the hyperparameters based on what leads to the highest number of successful attacks with the overall smallest difference in magnitude between clean and adversarial samples. The maximal number of iterations we use to search for a successful adversarial example is 100; the target t is chosen for each sample separately to be any random but different prediction (in comparison to the prediction on a clean sample). For C&W, we set the maximum number of iterations to 1000; the remaining hyper-parameters, i.e. clipping factor , update factor g and weight factor a, are tuned on the audio validation set (cf. Sect. 4.1). They are chosen as the setting that leads to the highest number of successful adversarial perturbations. The target t for each sample is, due to the binary nature of the SVDS, the class that is not the original prediction. For exact hyper-parameters, we refer to Tables 4 and 5 in Appendix.  1 Example explanations for each of the used standard explainers. Different colours visualise the weights/ attributions that different explainers assign to particular segments/pixels, with higher numbers indicating more, and lower numbers less ''important'' features for a prediction

Experimental setup
In this section, we describe the experimental setup, including the data we use and the classifiers we investigate in our experiments. We provide the code via Github 2 ; for further implementation details, we refer to Appendix C.

Data
To train the singing voice detection system (cf. [6]), we use the openly available Jamendo dataset [30], which consists of 93 songs resulting in roughly 6 hours of music. The training / validation / test split is proposed to contain 61 / 16 / 16 songs, respectively, with roughly the same proportion of annotated singing voice versus non-singing voice in all three splits. The audio files have a sampling rate of 44.1kHz, and are annotated with the presence (''sing'') or absence (''no sing'') of singing voice on a subsecond granularity [31]. Note that during training and test time, songs are not used as a whole, but instead split into a multitude of smaller excerpts (length % 1:6 seconds).
To extend these experiments to the image domain, we use the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) dataset [32]. The dataset contains 1,000 different classes and provides around 1.2 million images for training, as well as 50,000 images for validation (i.e. 50 per class). Subsequent experiments are performed on all validation images or a subset thereof 3 . Hyper-parameters are tuned on validation data as well, as we assume an attacker as well as people looking for explanations have access to the respective data.

Singing voice detection system
For experiments on audio data, we use the SVDS previously introduced by Schlüter and Lehner [18]. Data preprocessing, which consists of resampling and computing normalised magnitude Mel spectrograms, as well as the training procedure itself is done as proposed for the suggested Convolutional Neural Network (CNN) architecture.
To compute performance metrics of the SVDS, we use binary predictions of the network, which are obtained by application of a median filter followed by thresholding (cf. [18]). The classification error (%) on the test data is 11:54 AE 0:96 (given as mean ± standard deviation over 5 runs with different random initialisations). Recall and specificity are 89:61 AE 1:71 and 87:46 AE 1:00, respectively.
The top-1 accuracy of each pre-trained model on the validation data of ImageNet is shown in the second column in Table 1, as well as the accuracy of the classifiers after an attack (column 3). For more information regarding the networks we used, we refer to our implementation details in Appendix (section C).

Evaluation of LIME in the audio domain
In our first set of experiments, we look at explanations of the predictions on singing voice data. We first compute adversarial excerpts and explanations for their predicted class. Each of the explanations consists of a list of 20 interpretable features (rectangular segments) and their corresponding weight, which is interpreted as its importance for making a prediction [17,19]. We also present a baseline for choosing candidates of ''most influential'' segments of an input.

Explaining predictions on adversarial examples
To evaluate whether LIME can successfully detect causes for a prediction, we use (imperceptible) adversarial perturbations as ground truth. The goal is to investigate whether the explanation can highlight the segments that are responsible for the adversarial (and new) prediction.

Using a fixed number of interpretable features
It is often claimed that a small number of interpretable features k are sufficient to explain a classifier's prediction [17]. In the particular case of our singing voice detector, a choice for k in existing work is 3 [19]. We use LIME to explain predictions after successful adversarial attacks where perturbations of the original signal could have appeared anywhere in the spectrogram and also on possibly more than 3 segments. We then perform another attack adding only those perturbation segments coinciding with the 3 segments highlighted by LIME and evaluate how well they are able to change a prediction. We argue that if these highlighted segments actually explain a particular prediction, we can expect using this subset of perturbed segments to achieve the same change in prediction as the full perturbation did. However, using k ¼ 3 changes only 32% (''sing'' ! ''no sing'') and 41% (''no sing'' ! ''sing'') of the predictions. Subsequently, we call this metric the label-flip rate.

Varying the number of interpretable features
As fixing k ¼ 3 could also simply be a suboptimal choice for k, we perform this analysis for all k 2 f1; :::; 20g. Additionally, we select k perturbation segments based on their magnitude (i.e. the squared L2 norm) as a baseline. The according label-flip rates are shown in Fig. 2.
Almost regardless of the value for k, the explanations provided by LIME do not reach the performance of the baseline. For LIME-chosen segments, it requires almost all 20 segments to achieve label-flip rates close to 100%, whereas for the baseline, this is the case with a fraction of the number of segments.

Explaining ''localised'' perturbations
As it is often sufficient to add even small numbers of selected segments of a perturbation and still achieve new classifications, we conduct an additional experiment on how different explainers can recover changed segments of such refined or partial adversaries. We refine adversarial perturbations by splitting them into segments that correspond to the segments we use for LIME (i.e. rectangles). We then choose the k segments with the highest magnitude and use them to perturb the original (clean) data. In what follows, we show results on adversarial subsets of the data, namely samples for which adding k 2 f1; 3; 5g adversarial segments change ''sing'' to ''no sing'' or vice versa. for LIME compared to normbased baseline. The x-axis shows number of perturbed segments that are added to clean input Fig. 3 Relative number of segments correctly recovered by LIME, after adding only k perturbed segments to each sample. We look at 52 / 146 / 215 adversaries in total for k ¼ 1=3=5 We use LIME to compute explanations for this set of refined adversaries and set the number of interpretable features k to be equal to the number of perturbed segments in a sample (i.e. 1, 3 or 5). Since the segments of the explainer and the partial perturbations align, we can examine how often LIME is able to recover the correct segments that are perturbed, and hence are responsible for the new prediction. The result of this experiment is shown in Fig. 3; even for k ¼ 1, LIME shows the correct segment as explanation in less than half of the cases. Also for k ¼ 3 or k ¼ 5, the explainer rarely finds all modified segments, Fig. 4 First row shows the original input image, the raw IG attributions, and the most relevant pixels in binarised form (as used in [14]). The second and third rows show the proposed techniques for aggregating pixel-wise attributions using rectangular and SLIC segments, respectively. Colours again show importance of segments, with higher values denoting higher importance and in some settings none of the segments responsible for a prediction.

Evaluation of explanations in the image domain
We extend our previous experiments by analysing image classification. We once more split all perturbations into the segments (here: image subparts) that the explainers work with, i.e. now either rectangles 4 or SLIC segments. We are using 4 Â 4 ¼ 16 rectangular segments or set n_seg-ments=16 for the SLIC experiments 5 . To ensure a fair comparison between LIME and other explainers, we propose to aggregate pixel-wise attributions using the same segments that LIME uses. The aggregation procedure is explained in more detail subsequently, followed by a set of experiments for evaluating different explanations in the image domain. As a baseline for our explainers in the following experiments, we again use the selection of most important segments as the segments exhibiting the highest norm. Due to the huge computational cost for computing explanations, the experiments in Sects. 6.2 and 6.3 are performed on 10% of the validation dataset.

Aggregating pixel-wise attributions into segments
In order to compare Sal, GBP and LIME, Göpfert et al. [14] treat the segment weights computed by LIME as pixelwise attributions. They fix the number of ''most influential pixels'' that contribute to an explanations for all explainers, and compute the overlap between the pixel-wise attributions and the ground truth, i.e. all pixels in a segment that were adversarially perturbed. This might be unfair toward gradient-based explainers since pixel-wise attributions are often spread out over the whole image yet still could be concentrated around the most relevant regions. The localised perturbations of Göpfert et al. [14], however, only cover constrained parts of the image.
We focus on highlighted regions as opposed to highlighted pixels. Even publications about saliency-based attribution maps often talk about finding or highlighting regions [23], highlighting areas or regions [38], or ''descriptive image regions'' [16] rather than pixels. We further motivate this approach by the fact that attribution maps look like they are highlighting regions, e.g. CAM [39], Grad-CAM [40], Occlusion [41], and Deconvolution [41]. Recently, there have also been attempts at smoothing noisy attribution maps in order to highlight meaningful regions instead of seemingly random pixels [23].
Due to the fact that the goal of many of these explainers is highlighting regions rather than single pixels, it seems fair to also evaluate their capability in doing so. We propose aggregating pixel-wise attributions per segment to receive one importance value per segment, as we obtain for LIME. The only other work we are aware of who aggregate pixel-wise attributions into segments (or ''patches'') [42] uses an aggregation method called ''total patch saliency'', without specifying how pixels are actually aggregated.
We compare three approaches: • mean: average over raw pixel-wise attributions per segment; • mean_abs: average over the absolute pixel-wise attributions per segment; • mean_pos: average over the positive pixel-wise attributions per segment.
An example for each of the aggregation approaches, along with an example of how Göpfert et al. [14] post-process attribution maps, is shown in Fig. 4.
In preliminary experiments, mean_abs led to the highest label-flip rates (cf. Sect. 5.1), which is why we report results using this approach below.

Using a fixed number of interpretable features
Similar to the experiments in Sect. 5.1, we compute explanations for all adversarial examples, this time for 13 different explainers and two types of segmentation: rectangles and SLIC. For each explainer and segmentation, we fix k ¼ 3 and add segments of perturbations coinciding with the most influential segments to the original input and count how often the label is changed to compute the labelflip rate. Results for explanations using SLIC segmentation are found in Tables 2 and 3. The label-flip rates for the rectangular segments are summarised in Appendix in Tables 7 and 8. The results show the following: (i) Explanations in the image domain appear more meaningful when using SLIC segments compared to rectangular segments for k ¼ 3. (ii) When using rectangular segments, all standard explainers perform similarly across models, with GBP slightly outperforming the others for AlexNet, VGG16, and ResNet-50; and LIME slightly outperforming the others for DenseNet and Inception v3. (iii) Explainers generally perform best for AlexNet, followed by VGG16, and worst for DenseNet161. (iv) When using SLIC segments, LIME outperforms all other explainers by a large margin. The ranking of LIME, GBP, and Sal is consistent with the results in [14] (who analysed Inception v3) for AlexNet, VGG16, ResNet-50, and Inception v3.
(v) Smoothing does not always help and in many cases it even decreases the performance. There is a noticeable pattern, however, when smoothing helps. For ResNet-50, DenseNet161, and Inception v3, smoothing increasesacross all ensemble methods-the label-flip rate (i.e. the relative number of times parts of a perturbation change the original prediction of an input) for the setting with SLIC segments. (vi) In most cases, 3 segments are not sufficient to ''explain'' a prediction.

Varying the number of interpretable features
We also look at the label-flip rate of each explainer when increasing k and analyse the findings from the previous section in more detail. For rectangular segments, we perform this analysis for all k 2 f1; :::; 16g. Due to the varying number of SLIC segments (cf. Table 6), we follow a different approach: we iteratively add ½20; 30; :::; 100% of all segments. Based on the findings in the previous experiment, we particularly want to investigate the following: Rectangles vs. SLIC: For this experiment, we also add rectangular segments in percentage steps. Figure 5 shows the results for LIME, IG and Sal on ResNet-50. LIME works better when SLIC segments are used, while the other explainers work better with rectangles. This finding holds across all models. Standard explainers vs. baseline: Figure 6 shows that in contrast with the audio domain, all explainers are at least on par with our proposed magnitude baseline. For AlexNet and Inception v3, all explainers slightly outperform the baseline, for DenseNet only LIME manages to do so. Comparing different architectures: In Fig. 7, the performance of standard explainers for different model architectures is compared. As already suggested by the results in the previous section, it seems ''easier'' to explain predictions made by AlexNet as indicated by the higher label-flip rate across all explainers. VGG16 follows AlexNet in all comparisons. For the rest of the models, it is not as clear, but in general explainers perform (slightly) better for ResNet-50 and worst for Inception v3 and DenseNet161.

Explaining ''localised'' perturbations
We also want to look at refined adversarial image perturbations. After splitting the perturbations into image segments corresponding to the segments used by our explainers, the k segments with the highest magnitude are chosen and used to perturb the original data. In Fig. 8, we show results on adversarial subsets of the data for which adding k 2 f1; 3; 5g adversarial segments change the original prediction of a network to any new classification 6 . As the results for the explainers vary only slightly for different settings, Fig. 8 shows exemplary results for Inception v3 on rectangular and SLIC segments (experiment performed on 10% of validation data). The x-axes show different explainers; the y-axes show, in different colours, the fraction of times a particular amount of segments (out of k) has been recovered by the explainer. For more experimental results, we refer to Appendix (Sect. E.1). Figure 8 shows that all explainers we look at struggle to recover perturbed segments. In the extreme case, i.e. when perturbing only a single segment (left-most plots), the majority of segments is not detected (darkest colour), only improving slightly if we look at SLIC segments. For increasing k, the explainers tend to detect 1-3 (out of 3 or 5) perturbed segments (medium colours). In the rarest cases, all segments are recovered correctly. The main observable trends are that SLIC segments tend to be easier to detect except in the case of AlexNet (results omitted here) and that standard explainers, i.e. LIME, IG, Sal, and GBP, usually slightly outperform their more advanced counterparts.

Discussion and conclusion
In this work, we (i) investigated the performance of different explanation methods on detecting non-obvious causes for a prediction by using adversarial perturbations as a ground truth and (ii) compared these explainers to a new non-random baseline. We tested explainers in two different domains, namely for audio (singing voice detector) and images (different ImageNet classifiers). To treat perturbation-and gradient-based explainers fairly, we proposed a pixel-wise attribution aggregation, where different granularities are overcome by mapping all attributions to interpretable segments.
Our overarching result is that explainers struggle-regardless of the particular domain, task or model architecture-to find the parts of inputs that are really relevant for a system's prediction. We were able to demonstrate this empirically via a number of carefully designed experiments providing numerous counterexamples to one of the central claims of the field of explanation: being able to ''explain[s] the predictions of any classifier in an interpretable and faithful manner'' [17]. Our counterexamples disprove such universal statements concerning explanation methods. In more detail, our results show that: • In the audio domain, the explanation method LIME is not able to recover perturbed segments in a satisfactory manner, with even the baseline performing better • In the image domain, all investigated explanation methods at least surpass the baseline, but nevertheless are again not able to recover perturbed segments correctly • In the image domain, explainers perform somewhat better in recognising perturbed segments when looking at ''simpler'' image classifier architectures • For images, LIME performs best among the explainers (particularly when using SLIC segments), while standard gradient-based explainers (e.g. IG, Sal, GBP) usually outperform their more advanced versions There are two major lessons from our work that could move the field of explainability forward: (i) In line with previous results [14], we recommend using ground-truth based on adversarial perturbations. This approach, and also more general dataset modifications (cf. [13]), will no longer rely on questionable ground-truth provided by human observers, avoid their confirmation bias (cf. [2]), and hence allow for a more objective comparison of methods. (ii) We recommend to always use and compare different explainers since our experiments indicate that methods perform differently depending on the application domain (cf. LIME performing below baseline for audio but clearly above for images). This is in line with previous results where explainers perform differently even for different datasets within the same domain [43].
Supplementary information Not applicable.

Appendix A Explanations
In this section, we show exemplary explanations for different ensemble explainers we compare in this work in Fig. 9. The figure shows an input image and how the nine different ensemble explainers would explain the (correct) prediction of the image.

B.1 Hyper-parameters
In this section, we list the exact hyper-parameters used to attack the image classifiers (Table 4) with PGD and the singing voice detector (Table 5) with the C&W-based attack. Fig. 9 Example explanations for each of the used ensemble explainers. Different colours depict the weights/attributions that different explainers assign to particular segments/pixels, with higher numbers indicating more, and lower numbers less ''important'' features for a prediction

E.2 Explaining ''localised'' perturbations
In what follows, results omitted in Sect. 6.4 are shown. We show an excerpt thereof, as the results vary only slightly. Figure 13 shows results when trying to recover a fixed number of perturbed segments with different explainers for two architectures for rectangular segments. Figure 14 shows the same for SLIC segments. All plots show the results when perturbing k 2 f1; 3; 5g segments; the x-axes show different explainers, the y-axes their success-rate in recovering a certain number out of all changed segments.  Fig. 13 Percentage of correctly recovered k segments after adding k 2 f1; 3; 5g rectangular segments of a perturbation, for different explainers Funding Open access funding provided by Austrian Science Fund (FWF).

Declarations
Conflict of interest The authors of the manuscript confirm that there are no conflict of interest to declare.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons. org/licenses/by/4.0/.