A concept bottleneck model is an image classifier forced to first predict a handful of human-readable attributes ("red wing", "hooked beak") and then make its final decision only from those attributes. Its appeal is that a person can read the attributes as an explanation and correct a wrong one to fix the answer. In practice the explanation is often a cover story. The continuous attribute scores secretly carry extra information about the label and about irrelevant image content, and the downstream classifier learns to read those hidden signals. This is called leakage, and it makes human corrections unreliable. Separately, standard image codecs such as JPEG, JPEG2000 and VVC are rigid and content-blind: they discard fine, low-energy detail by frequency band rather than by meaning. That rigidity is usually counted as their main weakness, but it also makes them unusually hard for adversarial noise to exploit.
This note proposes to test one hypothesis: leakage in concept models and their adversarial fragility travel on the same carrier, the fine-scale image content that a hand-crafted codec throws away first. If that is true, a fixed codec placed in front of a concept model is a single cheap layer that makes the explanation more honest and the system harder to attack. Its failure modes would also differ from those of model-internal fixes, which is what defense-in-depth asks of a layer. If it is false, the codec only removes concept-relevant detail, and the experiments will show that cleanly. The contribution is the measurement that settles the question, together with a layered system built on whichever answer holds.
Core Claim
The claim has three parts, and each can fail on its own. First, a measurable share of concept leakage is carried by high spatial frequencies. Leakage here means two things: information about the label that the soft concept scores carry beyond what their thresholded 0/1 versions carry, and information about nuisance factors such as background. Second, removing that content with a fixed-transform codec at a matched bitrate reduces leakage and makes interventions more reliable, more than a content-blind low-pass blur at matched distortion does. Third, the same codec layer reduces the success of attacks that try to rewrite the explanation (flip concepts while keeping the label) under non-adaptive and cheap non-gradient attacks. We will report separately how much of that protection survives an attacker who differentiates through a smooth approximation of the codec. We expect the faithfulness benefit to be the more durable part. Leakage is not adversarial, so a defense against it cannot be "broken" by an adaptive attacker in the same way.
Relation to Prior Work
Two papers motivate the leakage side. Addressing Leakage in Concept Bottleneck Models shows that soft concept probabilities act as side channels that encode label information. It fixes this inside the model with an autoregressive hard concept predictor plus an explicit, regularized side channel. There Was Never a Bottleneck in Concept Bottleneck Models argues that predicting a concept from a latent never stops that latent from encoding extra content. It adds a variational penalty on the conditional mutual information between each concept latent and the input. Both act on learned representations and say nothing about which part of the image the leaked information comes from. We ask that upstream, physical question, and we treat their methods as the strongest competitors rather than as things to replace.
The robustness side is motivated and constrained by compression work. Attack and Defense Analysis of Learned Image Compression reports that learned codecs collapse under gradient attacks: a 61.55% PSNR drop, or a 19.15x bitrate increase. It also reports that VVC's fixed frequency transforms are far more resilient. That is why we use hand-crafted codecs as the sanitizing layer and treat learned codecs as a fragile comparison point. A Training-Free Defense Framework for Robust Learned Image Compression shows that random spatial transforms break the pixel alignment of adversarial noise without retraining. It supplies a second, diverse input layer, but it evaluates only reconstruction, never downstream explanations. Perceptual Quality Study on Deep Learning Based Image Compression documents that hand-crafted codecs produce blocking and ringing at low rates. That warns us that fine-grained concepts may be destroyed, so we must measure damage per concept rather than on average.
The closest existing combination is Neural Image Compression and Explanation (NICE). NICE learns a sparse pixel saliency mask and compresses the background more heavily, reaching 1.6x compression on Caltech256 for a 3.35% accuracy drop. Our proposal differs in direction and object. NICE uses an explanation to drive compression. We use a fixed, unlearned compression to test and improve the faithfulness of concept explanations, and we add an adversary. NICE's masks were also never checked under retraining. A Benchmark for Interpretability Methods in Deep Neural Networks (ROAR) shows that such checks matter: removing 90% of ImageNet pixels at random still leaves a retrained ResNet-50 at 63.53%. That redundancy is exactly why every faithfulness claim here is measured with retrained concept heads, not by masking at test time.
On the systems side, Position: “Real Attackers Don’t Compute Gradients” argues that real attackers use cheap transforms and that evaluations should cover pipelines, not isolated networks. Synergistic Security for the Industrial Internet of Things shows formally that combining diverse mechanisms beats investing in one. Together they set our evaluation standard: the codec layer is worth adding only if its failures are not the same as the model-internal regularizer's failures.
Two papers set the black-box regime. Concept Bottleneck Models supplies the base architecture. Concepts from Representations (PCBM-ReD) retrofits concept heads onto frozen vision backbones with a 0.41% average accuracy drop. The post-hoc setting matters because a codec layer is one of the few interventions available when the backbone cannot be retrained.
Research Program
The first stage is a mechanism test that uses no codec at all. We train concept models on inputs restricted to individual bands of the discrete cosine transform: the 8x8 block frequency decomposition JPEG itself uses. For each band configuration we measure two kinds of leakage. Label leakage is the gap in label accuracy between a probe reading the soft concept vector and a probe reading its thresholded version. Nuisance leakage is the accuracy of a probe that predicts a known nuisance factor from the concept vector. On Shapes3D and MPI3D the generating factors are known exactly, so one factor can be held out as a pure nuisance. On CUB + Places the nuisance is the pasted background scene. The output is a spectral profile of leakage, and the hypothesis lives or dies here. If leakage sits in the same low bands as concept evidence, a frequency-based codec cannot separate them, and the program narrows to a robustness study.
The second stage inserts real codecs (JPEG, JPEG2000, VVC intra) at a sweep of bitrates in front of three model families. These are trained concept models, the two leakage-corrected models, and a post-hoc concept head on a frozen CLIP-style backbone. Every concept head is retrained or fine-tuned on codec-processed images, in the spirit of ROAR. Leakage, concept accuracy, task accuracy and intervention curves are reported against bits per pixel. Because traditional codecs are content-blind, we also measure damage per concept, binned by how fine-scale the concept's evidence is (wing pattern versus overall size). This turns the codec's known limitation into a reported quantity rather than a hidden one.
The third stage is adversarial. We attack the full pipeline with three threat models. The first is non-adaptive gradient attacks crafted on the uncompressed model. The second is cheap non-gradient manipulations of the kind Apruzzese et al. describe. The third is adaptive attacks that differentiate through a smooth approximation of the codec. Each attacker pursues two goals: change the label, or change the explanation while preserving the label. The second goal directly targets the gap between a plausible explanation and a faithful one. We then stack the random-transform layer from the two-way compression work and the leakage regularizer, and check whether the layers' failures overlap.
Baselines
The naive reference point is the undefended joint concept model on raw images. On CUB-200-2011, Concept Bottleneck Models reports 19.9% error for the joint model against 17.5% for a standard black box. On CUB + Places with shifted backgrounds, it reports roughly 48–49% error for the concept model against 62.7% for the black box. Any codec configuration that pushes clean CUB error well above 19.9% has paid too much.
The strongest existing competitors on leakage are the two model-internal fixes. The autoregressive hard predictor of Addressing Leakage in Concept Bottleneck Models reports 81.4% concept accuracy on CUB, against 48.1% for soft independent models, with task accuracy within about 1% of leaky models. The minimal concept bottleneck of There Was Never a Bottleneck in Concept Bottleneck Models is the most direct competitor, because it attacks exactly the nuisance information we target. For its CUB, Shapes3D or CUB + Places leakage figures, no reported result found on Lacuna. We will therefore reimplement it and report our own numbers.
Fairness means every model sees the same images and the same backbone. The regularizer's weight and the codec's bitrate each get the same number of tuning runs, selected on a validation split by the same criterion: lowest leakage subject to task accuracy within 1% of the joint model. Beating these baselines means one of two outcomes. Either the codec lowers nuisance and label leakage below the regularizer at equal task accuracy, or codec plus regularizer lowers leakage below the regularizer alone. Intervention curves must also be at least as monotone.
For robustness, the strongest competitor is PGD adversarial training of the concept model, given the same training-compute budget as our full stack. For the compression layer itself, Attack and Defense Analysis of Learned Image Compression reports a 95.52% reduction in rate-distortion cost under attack after adversarial training. That figure concerns reconstruction, not concepts. For attacks on concept explanations, no reported result found on Lacuna.
The key ablation replaces the codec with a Gaussian low-pass blur and with JPEG at a matched distortion level, so the comparison asks whether quantization structure matters or merely smoothing. A learned hyperprior codec at matched bitrate serves as the fragile comparison point.
Why not just…?
Why not just use the information-bottleneck regularizer and skip the codec? This is the real threat. If the minimal concept bottleneck already removes nuisance information, the codec may add nothing for faithfulness. The second stage settles this directly by comparing regularizer alone, codec alone, and both, at matched task accuracy. If the combination gives no leakage reduction beyond the regularizer, we will say so and drop the faithfulness claim. The codec would then survive only if it adds robustness the regularizer lacks, since the regularizer is a learned component that gradient attacks can reach.
Why not just put JPEG at a fixed quality in front of the classifier, as an old input defense? That is part of what we test, not an alternative to it. The open question is whether it changes explanation faithfulness, which no prior work in the file measures. Its well-known weakness against attackers who differentiate through a smooth approximation is exactly why we report adaptive attacks separately. We also make no robustness claim that rests only on non-adaptive numbers. The per-concept damage analysis addresses the further worry, raised by the artifacts documented in Perceptual Quality Study on Deep Learning Based Image Compression, that fixed quality destroys fine-grained concepts.
Why not just blur the image? A blur is the simplest frequency filter. If it matches the codec at equal distortion, the finding is "high frequencies carry leakage," and the codec is incidental. That would still be a useful result, but a smaller one. The matched-distortion ablation is designed to tell these apart. Block quantization discards low-energy coefficients adaptively within each block, whereas a blur discards them uniformly. We expect this to spare strong fine edges that carry concept evidence.
Why not just adversarially train the concept model? Adversarial training hardens one layer. The compression literature reports that it degrades clean performance and is expensive, and Laszka et al. show that single-mechanism investment hits diminishing returns. More importantly, adversarial training does not target leakage at all. We include it at equal compute, and we test whether codec plus adversarial training fails on different inputs than either one alone. Overlapping failure sets would undercut the defense-in-depth argument.
Decisive Checks
The first-stage spectral profile is the go/no-go test. Leakage must concentrate in bands that carry little concept accuracy on Shapes3D and on CUB + Places. If it does not, the faithfulness thesis is rejected before any codec is trained.
Second, at a bitrate where clean CUB error stays within one point of the joint model's 19.9%, the codec must reduce background-predictive nuisance leakage on CUB + Places. It must also make intervention curves more monotone, and do both better than matched blur.
Third, on the explanation-rewriting attack, the codec layer stacked with the regularizer must have lower success than either alone under non-adaptive and cheap attacks. Under adaptive attacks we report honestly whatever remains.
Finally, all of the above must replicate with a frozen black-box backbone and a post-hoc concept head. That is the regime where an input-side layer is the only practical lever.
References
- Addressing Leakage in Concept Bottleneck Models
- There Was Never a Bottleneck in Concept Bottleneck Models
- Attack and Defense Analysis of Learned Image Compression
- A Training-Free Defense Framework for Robust Learned Image Compression
- Perceptual Quality Study on Deep Learning Based Image Compression
- Neural Image Compression and Explanation
- A Benchmark for Interpretability Methods in Deep Neural Networks
- Position: “Real Attackers Don’t Compute Gradients”: Bridging the Gap Between Adversarial ML Research and Practice
- Synergistic Security for the Industrial Internet of Things: Integrating Redundancy, Diversity, and Hardening
- Concept Bottleneck Models
- Concepts from Representations: Post-hoc Concept Bottleneck Models via Sparse Decomposition of Visual Representations