Keep the Simplex, Spare the Spread: Separating Two Kinds of Collapse in Adversarial Training

When a classifier is trained to resist adversarial examples (tiny, deliberately chosen input changes that flip its prediction), its accuracy on ordinary, unperturbed images usually drops. Recent work offers a geometric clue to why. Late in training, a network's internal features tend to settle into a very regular shape called "neural collapse": every image of a class maps to nearly the same point, and the class centers spread out as far as possible from one another, like the corners of a regular triangle or tetrahedron (a "simplex"). Standard adversarial training strengthens this collapse. TRADES, a method known for a better balance between clean and robust accuracy, does not produce it at all. We propose that the "collapse" adversarial training induces is really two different things. One is pushing class centers onto a maximally separated simplex, which is plausibly what buys robustness. The other is squeezing out all variation within each class, which is plausibly what costs clean accuracy. We test this claim by controlling the two separately: we fix the class centers to a simplex and actively keep within-class spread during adversarial training.

If the claim holds, part of the robustness-accuracy trade-off comes from a removable side effect of how robust features are arranged, not from robustness itself. That would give a cheap, architecture-level fix that stacks on top of existing adversarial training. If it fails, the experiments still settle a question that On the Robustness of Neural Collapse and the Neural Collapse of Robustness explicitly leaves open: whether TRADES's lack of collapse actually causes its better clean accuracy.

Core Claim

Neural collapse has separable components. NC1 is within-class variability collapse. NC2 is the class means forming a simplex equiangular tight frame (a simplex ETF, meaning equal-length, equal-angle class vectors). Both were identified by Prevalence of Neural Collapse during the terminal phase of deep learning training. Our hypothesis is that under adversarial training, the drop in clean accuracy is driven mainly by NC1, while robustness is supported mainly by NC2.

Su et al. found that adversarial training produces "robust neural collapse," where clean and perturbed inputs land on the same simplex. They also found that TRADES reaches strong robustness without collapse. They describe "cluster leaping," in which an attack on a standard model moves a feature almost exactly onto another class's corner: about 11° from the target mean when class means sit about 96° apart. These findings motivate us, but they bundle NC1 and NC2 together. We split them.

The method uses two existing ingredients and applies them to a new problem. The first is the decoupled design of Controlling Neural Collapse Enhances Out-of-Distribution Detection and Transfer Learning. That design uses an encoder regularized with a differential-entropy term so its features stay spread out, followed by a frozen projector whose weights are fixed to a simplex ETF. Harun et al. showed that this split helps transfer and out-of-distribution detection, but they never tested adversarial attacks. The second ingredient is a standard adversarial objective (PGD adversarial training or TRADES) applied on top. The training loss is

$$\mathcal{L} = \mathcal{L}_{adv}\big(W_{ETF}\, g(h_\theta(x)),\, y\big) \;-\; \gamma\, \hat H\big(h_\theta(x)\big),$$

where $h_\theta$ is the encoder and $g$ is a small projector. $W_{ETF}$ is the frozen simplex head. $\mathcal{L}_{adv}$ is either the PGD loss or the TRADES loss. $\hat H$ is a nearest-neighbor estimate of the differential entropy of encoder features within a batch. $\gamma$ sets how strongly within-class spread is preserved.

Research Program

The core object is a two-by-two factorial design, run under both PGD adversarial training and TRADES. The first factor is the classifier head: either learned, or fixed to a simplex ETF. The second factor is the anti-collapse entropy term: present or absent. Every run logs the following quantities on clean and attacked test data, at the encoder output and at the projector output:

  • NC1, measured as within-class covariance relative to between-class covariance;
  • NC2, measured as the deviation of class means from a simplex ETF;
  • the angle between the clean simplex and the perturbed simplex;
  • the cluster-leaping angle under targeted PGD.

The analysis question is which collapse metric predicts the clean-accuracy loss across runs at matched robust accuracy. Our hypothesis predicts three outcomes. First, NC1 should explain most of that variance. Second, the fixed-ETF-plus-entropy cell should dominate the plain adversarial training cell. Third, TRADES should gain less from the intervention than plain adversarial training, because TRADES already avoids collapse. That last prediction directly tests the paradox Su et al. leave open.

A necessary side study calibrates the "free lunch" claims of passive simplex losses. Gaussian class-conditional simplex loss for accurate, adversarially robust deep classifier training and its journal version Beyond cross-entropy: learning highly separable feature distributions for robust and accurate classification report 97.06% clean accuracy and about 85% accuracy under a 5-step PGD attack on CIFAR-10, with no adversarial training. Max-Mahalanobis Linear Discriminant Analysis Networks reports 91.3% under FGSM at $\epsilon=0.04$. Neither was evaluated with AutoAttack. AutoAttack is the ensemble of Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks; it found lower robustness than originally reported in 48 of 49 published defenses. We re-run both passive losses under AutoAttack.

If they survive, the role of adversarial training in our design must be reconsidered. If they collapse, as we expect from gradient masking, that confirms the simplex alone does not confer robustness. In that case the simplex is a geometry to impose during adversarial training, not a replacement for it.

Baselines

The naive reference point is standard cross-entropy training with no defense. It sets the clean-accuracy ceiling and shows the full cluster-leaping vulnerability Su et al. describe.

The direct reference for the intervention is standard PGD adversarial training. On CIFAR-10 with WideResNet-32-10, TART: Boosting Clean Accuracy Through Tangent Direction Guided Adversarial Training reports 81.76% clean accuracy and 35.89% AutoAttack accuracy for it.

The strongest existing clean-accuracy-preserving method in the file is TART. TART estimates the data manifold in input space with an autoencoder and trains at zero perturbation size when the attack direction points off the manifold. It reports 83.73% clean accuracy and 36.58% AutoAttack accuracy (TART-AT), and 84.31% clean accuracy when combined with MART. It is the strongest competitor because it targets the same symptom, lost clean accuracy, with a different explanation: input-space manifold geometry instead of latent class geometry. We use the same backbone, the same $\ell_\infty$ budget of $\epsilon = 8/255$, the same number of epochs, and the same attack-step count. Each method receives an equal hyperparameter search budget over its trade-off knob: $\gamma$ for ours, the adaptive threshold for TART, and $\beta$ for TRADES. The proposal beats TART only if it exceeds 83.73% clean accuracy at AutoAttack accuracy of at least 36.58%. Because the two methods act on different spaces, we also test whether they stack.

TRADES (Theoretically Principled Trade-off between Robustness and Accuracy) adds a KL-divergence term that keeps clean and adversarial outputs close, with weight $\beta$. It reports 56.61% white-box robust accuracy on CIFAR-10. The file gives no clean accuracy for it, so we report our own sweep over $\beta$.

The key ablations are the two single-factor cells. The fixed ETF head without entropy tests whether imposing the simplex alone helps. The entropy term with a learned head tests whether spreading features alone helps. The design isolates the interaction only if both single factors fall short of the combined cell.

Why not just…?

Why not just tune TRADES's $\beta$? This is the real threat. TRADES already avoids collapse, so its Pareto frontier may already sit where ours does. We compare full clean-versus-AutoAttack frontiers, not single points. If TRADES's frontier matches the fixed-ETF-plus-entropy frontier under plain adversarial training, then the contribution narrows to an explanation of why TRADES works, not a better method. We would report it that way. The test that separates the two readings is whether adding our components to TRADES moves its frontier at all.

Why not just use a passive simplex loss like GCCS, which is cheaper? Its reported robustness comes from 5-step PGD and JSMA attacks. Croce and Hein showed that losses with non-standard output shapes are exactly the kind whose robustness evaporates under AutoAttack. Our side study measures this directly. We will not build on GCCS's numbers until they survive that evaluation.

Why not just reduce collapse with label smoothing, early stopping, or weaker weight decay? These change NC1 indirectly, and they also change calibration and margins. We include label-smoothed and early-stopped adversarial training as extra arms and plot each of them on the same NC1-versus-clean-accuracy axes. If they trace the same curve, the specific regularizer does not matter and the cheaper knob wins; that would still support the NC1 hypothesis. If they do not, the entropy term is doing something they cannot.

Why not just use TART? TART leaves latent geometry untouched, and its gain in AutoAttack accuracy is under one point. It also needs a pre-trained autoencoder and stored tangent matrices, which conflicts with online data augmentation. Our method costs one batch entropy estimate; Harun et al. report less than 0.3% overhead for their version. The stacking experiment shows whether the two methods address different causes.

Decisive Checks

The main experiments use CIFAR-10 with WideResNet-32-10 to match TART, and CIFAR-100, where the collapse geometry has 100 corners and the lower-rank alternatives of Neural collapse versus low-rank bias: Is deep neural collapse really optimal? may matter. ImageNette serves as a scale check matching Su et al. All robustness numbers use full AutoAttack at $\epsilon = 8/255$. Every attack must be able to target the frozen ETF head, and we add a check that more PGD steps keep lowering accuracy, to catch gradient masking.

Three results decide the claim:

  1. The factorial outcome. The combined cell must beat plain adversarial training on clean accuracy at matched AutoAttack accuracy, and neither single factor alone may do so.
  2. The NC1 regression. Across all runs, NC1 at the encoder must predict clean-accuracy loss better than NC2 does.
  3. The cluster-leaping angle. The combined model must keep perturbed features far from rival corners, so that preserved spread does not create new leaping paths.

Failing the first check while passing the second would mean NC1 explains the trade-off but our regularizer does not control it well. Failing the second would falsify the core claim.