Flattening the Wrong Answers: Complement-Entropy Regularization Against Catastrophic Overfitting in Single-Step Adversarial Training

Adversarial training makes image classifiers resistant to small, deliberately chosen input perturbations. It does this by training on worst-case perturbed inputs, which are usually found with 10 or more gradient steps per batch. That inner search multiplies training cost roughly tenfold. A cheaper variant uses a single gradient step (FGSM, the fast gradient sign method), but it often suffers catastrophic overfitting: partway through training the model becomes immune to the one-step attack yet drops to near-zero accuracy against multi-step attacks. This note proposes a loss-level fix built from an idea borrowed from machine unlearning. On the perturbed training example, push the probability the model assigns to the wrong classes toward a uniform spread, so that no single wrong class can become the easy escape route the one-step attack learns to exploit.

The claim is specific and falsifiable. Catastrophic overfitting is preceded by a measurable collapse in the entropy of the model's wrong-class distribution on single-step adversarial points. A confidence-gated penalty that keeps this distribution flat should prevent the collapse and yield robustness under AutoAttack at single-step cost. Because the penalty leaves the true-class term alone, it should also cost less clean accuracy than scalar-weighted alternatives. If this holds, one diagnostic quantity connects three problems: the unlearning community's preference for uniform targets over forced errors, the computational cost of adversarial training, and the robustness–accuracy trade-off.

Core Claim and Mechanism

Uniform-distribution targeting is the idea that a model should respond to information it should not rely on with maximal uncertainty rather than a confident wrong answer. The unlearning literature supplies both the motivation and the method. Forget by Uncertainty: Orthogonal Entropy Unlearning for Quantized Neural Networks shows that forcing wrong labels makes networks "misremember" a new spurious association, while maximizing output entropy avoids this. Yingdan Shi et al. show "fake forgetting", where the top-1 prediction changes but confidence still concentrates on a particular class. Our hypothesis is that the single-step attack's failure mode has the same structure. An FGSM step moves the input toward one dominant runner-up class. As training proceeds, the network learns a narrow, locally curved response in which that one direction stops being informative. The perturbed point's wrong-class mass concentrates, and the linear approximation the attack relies on breaks.

The method comes from Improving Adversarial Robustness via Guided Complement Entropy, which trains on clean data only. It flattens non-true-class probabilities, gated by a power of the true-class confidence so that flattening only begins once the correct class is learned. That paper reports lower clean error than cross-entropy on Tiny ImageNet (38.56% vs 39.54%). It also admits it "is not a total shield" against strong multi-step attacks and never runs AutoAttack. We move this term from the clean input to the single-step adversarial input and keep the standard adversarial cross-entropy. For a training pair $(x, y)$ with FGSM-perturbed input $x'$, model probabilities $p(\cdot \mid x')$, and $\hat p_c(x')$ the probabilities over the $K-1$ wrong classes renormalized to sum to one, the loss is

$$\mathcal{L} = \mathrm{CE}\big(p(\cdot\mid x'),\, y\big) \;-\; \beta\, p_y(x')^{\alpha}\, \frac{H\big(\hat p_c(x')\big)}{\log(K-1)},$$

where $H$ is Shannon entropy, $\alpha$ is the confidence gate (default $1/3$ following GCE), and $\beta$ sets the strength. The normalized complement entropy $H(\hat p_c)/\log(K-1)$, averaged over a held-out batch of FGSM points, is also our proposed early-warning statistic for catastrophic overfitting.

This objective differs from its neighbors in specific ways. TRADES (Theoretically Principled Trade-off between Robustness and Accuracy) matches the perturbed and clean predictions with a KL term, which says nothing about how wrong-class mass is arranged, and it needs a multi-step inner loop. Catastrophic Overfitting, Entropy Gap and Participation Ratio detects catastrophic overfitting from gradient statistics and fixes it by switching the perturbation norm between $\ell_2$ and $\ell_\infty$. It changes the attack, whereas we change the loss and monitor the output distribution. The two could compose, and we test whether they are redundant. AUTE (AUTE: Peer-Alignment and Self-Unlearning) "unlearns" over-confident adversarial points with negative loss weights inside a costly ensemble; it never targets a uniform distribution.

Research Program

The work has three stages. The first is diagnostic. We run FGSM and random-start FGSM adversarial training on PreAct ResNet-18 at $\epsilon = 8/255$ and $16/255$, where catastrophic overfitting is frequent, across many seeds. We log per epoch the complement entropy on FGSM points, the Entropy Gap and Participation Ratio from Mehouachi and Jabari, the gradient norm, and multi-step PGD accuracy. The question is whether complement-entropy collapse reliably precedes the PGD accuracy crash, and whether it precedes it earlier or more reliably than the gradient statistics. If it does not, the mechanistic story is wrong, and the intervention would at best work for other reasons.

The second stage is the intervention: the loss above, with the $\alpha$ gate, trained at single-step cost. The third stage separates real robustness from gradient masking, meaning a defense that only hides useful gradients from the attacker. This is the central risk of the proposal. Flattening output distributions is exactly the kind of change that can shrink gradients without widening margins. RobustBench documents defenses whose robustness was overestimated "sometimes by more than 10%" until AutoAttack was run. Every robustness claim here is therefore made under AutoAttack. It is also accompanied by the gradient-free Square attack, an adaptive attack that directly minimizes the true-class logit margin (so it ignores the entropy term), and an accuracy-versus-$\epsilon$ curve that must fall to zero.

Baselines

The naive reference point is standard training plus plain FGSM adversarial training with no safeguard. It establishes the catastrophic overfitting rate we are trying to eliminate and the cost floor. The strongest existing single-step competitor is the adaptive-norm method of Catastrophic Overfitting, Entropy Gap and Participation Ratio. The file reports it beats random-start FGSM by "approximately 3.3% in adversarial accuracy at $\epsilon = 6/255$" on ImageNet-1k, and it attacks the same failure through a different lever. We reimplement it from the paper's description, share our architecture, schedule, and data, and give it the same hyperparameter search budget as our $\beta$ and $\alpha$.

The accuracy-and-robustness ceiling we aim to approach is TRADES with a 10-step PGD inner loop. The file reports it reaches white-box robust accuracy "as high as 56.61%" on CIFAR-10. The file has no TRADES clean-accuracy figure and no TRADES AutoAttack figure, so we rerun it ourselves on the same architectures under AutoAttack. For CIFAR-100, the file's strongest AutoAttack reference is NNPRAT (Nearest Neighbor Projection Removal Adversarial Training) at "26.31%" AutoAttack robustness on WRN-34-10. We treat it as a full-cost upper reference rather than a direct competitor. Original GCE on clean data only is also run, because it is the paper we build on and the "no adversarial examples at all" extreme.

We would count the proposal as beating the field if it met three conditions at the same time. First, it drives the catastrophic overfitting rate at $\epsilon = 8/255$ and $16/255$ to zero across seeds. Second, it matches or exceeds the adaptive-norm method's AutoAttack accuracy at equal compute. Third, it closes most of the AutoAttack gap to our TRADES-PGD10 rerun, with equal or higher clean accuracy, at roughly one-fifth the training cost.

Two ablations isolate the key ingredient. The first applies uniform label smoothing to the adversarial point, which spreads mass over all classes including the true one. The second drops the confidence gate ($\alpha = 0$). If either matches the full method, the "complement" and "guided" parts are not doing the work.

Why not just…?

Why not just use label smoothing on the adversarial examples? This is the most serious alternative, because it also moves mass toward uniform. It differs in two ways. It pulls down the true class, and it applies the same pressure regardless of how well the class is learned. GCE's authors argue this creates a flat region that traps optimization. Jing Wu et al. and OEU show that fixed soft or random targets can themselves be memorized. Smoothing is also widely suspected of producing gradient masking rather than margin. None of this is decisive for the adversarial setting, so the label-smoothing ablation is run with the same tuning budget and the same masking checks. If it matches us under AutoAttack and Square, we would report that the effect is generic output flattening.

Why not just run random-start FGSM or single-step TRADES and tune $\lambda$? Mehouachi and Jabari's results show that random-start FGSM still overfits catastrophically at larger $\epsilon$. A scalar trade-off weight changes how much the robust term matters, not which failure mode it has. We include single-step TRADES with a swept $\lambda$ as a baseline. If it avoids catastrophic overfitting at $16/255$ with equal clean accuracy, our loss adds nothing.

Why not just use the adaptive-norm fix alone? It may be enough. Our diagnostic stage tests whether complement-entropy collapse and the Entropy Gap signal the same event. We also train the combination: if adding our term to the adaptive-norm method gives no gain, the two are redundant and the simpler published fix wins.

Why not skip adversarial examples and just use GCE on clean data? GCE's own authors say it does not withstand strong multi-step attacks. Its reported ~15% under the momentum iterative attack on CIFAR-10 is far below adversarially trained models. We rerun it under AutoAttack to show this directly.

Proposed Experiments

The main experiments use CIFAR-10 and CIFAR-100 with PreAct ResNet-18 and WRN-34-10 at $\ell_\infty$ radii $8/255$ and $16/255$, over at least five seeds. Metrics are clean accuracy, PGD-50 accuracy, AutoAttack accuracy, Square-attack accuracy, catastrophic overfitting rate, and GPU-hours. Tiny ImageNet tests whether the benefit grows with more classes, where the complement distribution has more room to concentrate. There, GCE's clean-error figure offers a clean-accuracy sanity check.

If CIFAR results hold, a ResNet-50 run on ImageNet-1k at $\epsilon = 4/255$ and $6/255$ compares directly against the adaptive-norm method's reported regime. Calibration is measured throughout, because RobustBench notes that robust models tend toward underconfidence. An output-flattening loss could worsen this, and that would be a real cost.

Three checks decide the proposal. First, complement-entropy collapse must precede the PGD crash in unregularized FGSM runs. Second, the regularized runs must show no catastrophic overfitting and an AutoAttack-versus-PGD gap no larger than TRADES's. Third, the label-smoothing and ungated ablations must fall short under the margin-based adaptive attack. Failing the second check on masking grounds would falsify the proposal even if PGD numbers look strong.