Image classifiers can be fooled by small, carefully built changes to their input, and such changes tend to push the image off the "natural data manifold." That manifold is the thin, structured region of pixel space where real images actually lie. The leading repair is purification. It means running the input through a generative model so that it is pulled back toward realistic images before classification. The best known version uses a diffusion model, a generator that learns to turn random noise into images step by step: the method adds noise to the input and then denoises it. Purification has a known weakness. How much noise to add is a dilemma with no good answer. Too little leaves the attack intact. Too much erases the features that determine the label, so a cat can come back as a dog. Once careful attackers compute correct gradients through the purifier, published robustness largely disappears.
We propose to stop trying to resolve that dilemma and to use it as a reason to abstain. The system classifies the raw input with an adversarially trained network, then classifies several purified copies made at different noise levels. It answers only when all of these views agree with high confidence, and otherwise it defers the input. The bet is that an attacker who wants a wrong answer accepted must fool views that fail in different ways, all at once and toward the same wrong label. That joint task should be harder than fooling any single purified copy or any single classifier. The question we test is whether this gives lower error on accepted inputs under a fully adaptive attack, at a fixed coverage on clean data, than the cheap alternatives.
Core Claim
Selective prediction, meaning classification with an explicit option to abstain, makes a recurring finding: confidence scores computed from the classifier's own outputs rank in-distribution errors well. Plain thresholding of the highest softmax probability reaches 1% error at 78% coverage on CIFAR-10 in Selective Classification for Deep Neural Networks. The same scores fail to reject inputs that have left the training distribution, as documented across 44 datasets in Deep Neural Network Benchmarks for Selective Classification.
Purification research makes the opposite trade. It actively moves inputs back toward the manifold, but it never abstains, so it must output a label even when the purification has visibly failed. Our claim is that the disagreement between purification at different noise levels and the unpurified, adversarially trained view is a selection signal. A small noise level keeps the label but may keep the attack. A large noise level removes the attack but may remove the label. An adversarially trained raw classifier is robust inside its training norm ball but fails off it. An input on which all three agree is one where the noise-level trade-off did not bite, and we accept it. An input on which they disagree is one we should not answer.
Concretely, let $v_k(x)$, for $k = 0,\dots,K$, be the class-probability vectors from the views. Here $v_0$ is the adversarially trained classifier on the raw input, and $v_k$ for $k \ge 1$ is the same classifier applied to diffusion-purified copies at noise levels $t_1 < \dots < t_K$, each averaged over a few stochastic samples. Let $\hat y(x)$ be the raw prediction. The confidence score is
$$\kappa(x) = \min_{k} \, v_k(x)_{\hat y(x)},$$
and we accept, setting $g(x)=1$, exactly when $\kappa(x) \ge \tau$. The threshold $\tau$ is set by the binomial-bound procedure of Selective Classification for Deep Neural Networks, which guarantees a target error rate on clean data with high probability. Under attack the guarantee does not transfer, so worst-case accepted error is measured empirically rather than claimed.
Relation to Prior Work
The closest work is Revisiting Adversarial Robustness of Classifiers With a Reject Option. It trains a classifier together with a rejection detector that reads the classifier's logits. It also defines "robust error with rejection": small perturbations must be accepted and classified correctly, while larger ones may be rejected but must never be misclassified. We adopt that metric. We differ in what drives rejection. Their detector reads the logits of a single network and is shaped by fixed norm balls. Ours reads the agreement between different reconstructions of the input, which tracks whether purification preserved its meaning, and it needs no rejection training.
Two purification papers supply the components and also motivate the gap. Diffusion Models for Adversarial Purification provides the diffusion purifier and the noise-level dilemma. Adversarial Purification with the Manifold Hypothesis observes that the variational autoencoder's likelihood bound drops under attack. It uses that drop only as an objective to climb back toward the manifold, taking 32 or more gradient steps at about 102× forward-pass cost, and never as a reason to abstain. Enhancing the Adversarial Robustness via Manifold Projection places an autoencoder projection inside adversarial training. It always outputs a label, and its authors acknowledge that the projector itself can be attacked.
Two papers constrain the design. DiffBreak: Is Diffusion-Based Purification Robust? shows that accurate gradients collapse diffusion purification to 17.19% robust accuracy on CIFAR-10 and 12.11% on ImageNet. It also shows that low-frequency attacks reach about 0% even with majority voting over stochastic passes. This is direct evidence that disagreement across random samples at one noise level is a weak signal, because the attacker can make the samples agree on the wrong label. That is why our views vary the noise level and include a non-purified, adversarially trained view, rather than only resampling.
When Does Diffusion Purification Amplify Perturbations? shows that the gap between clean and attacked denoising paths rises and falls non-monotonically. Early, high-noise steps are more than twice as sensitive as late ones. This is a reason to expect different noise levels to fail differently. It also warns that the highest-noise views may be the easiest place for an attacker to inject effect.
Research Program
We build the view-agreement selective classifier on CIFAR-10, CIFAR-100, and a 1,000-class ImageNet evaluation subset. Each uses an adversarially trained base classifier and the pretrained diffusion purifiers used by DiffPure. Noise levels are chosen on clean validation data so that their errors are as decorrelated as possible, with $K \in \{1,2,3\}$ to measure what each extra view buys.
The attack is the core of the evaluation. We attack the whole pipeline with exact gradients through the reverse process, following DiffBreak's gradient method, with averaging over the purifier's randomness. Three attack objectives are used:
- An acceptance attack that maximizes $\min_k \log v_k(x)_{y'}$ for a wrong label $y'$, with the threshold term included so the attacker optimizes for being accepted.
- DiffBreak's low-frequency attack, retargeted to the same joint objective.
- A rejection-forcing attack that tries to push clean inputs below $\tau$ inside a small budget. This is the failure mode the SATR metric penalizes.
Budgets follow the $\ell_\infty$ $8/255$ protocol used in the amplification study, with a smaller accept-radius for the rejection metric.
We also test a two-stage version in which inputs with very high raw confidence skip purification, which addresses the latency problem. This version is evaluated under an attacker who knows about the shortcut, since the shortcut is the obvious target.
Baselines
The naive reference point is the adversarially trained classifier with softmax-threshold abstention, with no purification. It is cheap, and its errors inside the training ball are already bounded. If the multi-view score cannot beat it at matched clean coverage, the idea fails. We train it with the same architecture and data and set its threshold with the same binomial procedure. The file has no reported number for its accepted error under attack, so we measure it. The training-time variant from Towards Better Selective Classification, which uses self-adaptive training plus an entropy penalty with softmax-threshold abstention, gets the same treatment as a stronger version of this baseline.
The strongest existing purification baseline is DiffPure with majority voting. The comparison point is the 17.19% CIFAR-10 and 12.11% ImageNet robust accuracy reported in DiffBreak. We turn it into a selective method by abstaining when the vote margin is low. This is the key ablation: stochastic disagreement at a single noise level, with no raw view. Beating it means lower accepted error under the joint adaptive attack at the same clean coverage and the same number of diffusion passes.
The strongest rejection baseline is SATR from Revisiting Adversarial Robustness of Classifiers With a Reject Option. No reported result for SATR's robust error with rejection was found on Lacuna, so we retrain it on our backbone and evaluate it with the same metric and the same attacks.
A compute-matched deep ensemble of adversarially trained networks with softmax-threshold abstention gets the same inference FLOPs as our views. Ensembles are the most consistent selective classifiers in Deep Neural Network Benchmarks for Selective Classification, with relative error as low as about 0.117. This baseline tests whether our gain is just extra compute.
A density baseline thresholds the likelihood bound of a jointly trained variational autoencoder, in the style of Adversarial Purification with the Manifold Hypothesis. It is attacked jointly with the classifier.
All baselines get equal hyperparameter-search budgets on clean validation data only. All are attacked with the same iteration counts and the same number of restarts.
Why not just…?
Why not just threshold the softmax of an adversarially trained model? This is the real threat, because inside the training ball adversarial training already limits confident errors. Our expected advantage lies outside that ball, in low-frequency perturbations and in rejecting failed purification. That is where purification has been shown to add robustness to unseen threats: DiffPure reports +36% over adversarial training on spatial transformations. If accepted error under the low-frequency and acceptance attacks is no better than this baseline at matched coverage, we will report the multi-view score as unnecessary.
Why not just abstain when stochastic purification passes disagree? DiffBreak's low-frequency attack drives majority-voted purification to about 0% robust accuracy. That suggests the attacker can make random passes agree on a wrong label, so resampling alone gives a weak selection signal. The single-noise-level ablation tests this directly. If it matches the full method, then the raw adversarially trained view and the varied noise levels are not doing the work.
Why not just abstain on low likelihood under a generative model? Likelihood measures whether an input is off the manifold, not whether its label survived purification. It is also a differentiable target the attacker can raise directly. The density baseline, attacked jointly, shows whether this matters in practice. If it holds up comparably at a fraction of the cost, that is a cheaper result and we will say so.
Decisive Checks
The primary result is a risk-coverage curve under the joint adaptive attack: accepted error on attacked inputs plotted against coverage on clean inputs, alongside robust error with rejection. It is reported on CIFAR-10, CIFAR-100, and the ImageNet subset.
Three checks decide the outcome. First, the method must beat both the adversarially trained softmax-threshold baseline and the compute-matched ensemble at clean coverage of 80% or more. Second, its advantage over single-noise-level vote disagreement must persist under the low-frequency attack. Third, the rejection-forcing attack must not drive clean coverage much below that of the baselines.
We also measure per-class abstention rates. The aim is to learn whether purification reduces or worsens the rejection bias against minority classes reported in the selective-classification benchmark. The two-stage shortcut version is reported separately with its latency and its error under a shortcut-aware attacker.
References
- Selective Classification for Deep Neural Networks
- Deep Neural Network Benchmarks for Selective Classification
- Revisiting Adversarial Robustness of Classifiers With a Reject Option
- Diffusion Models for Adversarial Purification
- Adversarial Purification with the Manifold Hypothesis
- Enhancing the Adversarial Robustness via Manifold Projection
- DiffBreak: Is Diffusion-Based Purification Robust?
- When Does Diffusion Purification Amplify Perturbations?
- Towards Better Selective Classification