Adversarially trained image classifiers, meaning models trained on worst-case perturbed inputs so that small pixel changes cannot flip their predictions, pay for that robustness with clean accuracy. The standard remedy is TRADES. It adds a penalty that keeps the model's output on a perturbed input close to its output on the clean input, and one weight trades clean accuracy against robustness. That penalty is blunt. It never asks how the attack moves the network's internal features. It only asks that the final prediction not change. A network can satisfy it by making every feature direction insensitive, and the cheapest way to do that is to discard fine-grained, predictive features wholesale. This proposal targets that mechanism of clean accuracy loss.
The idea is to use the nuclear norm, the sum of a matrix's singular values and the standard convex stand-in for matrix rank, on a quantity nobody has regularized before. That quantity is the batch matrix of feature displacements caused by the attack, the adversarial features minus the clean features. Penalizing the rank of this matrix, rather than its total energy, asks the network to let attacks move features only within a small shared subspace. The classifier head can then learn to ignore that subspace while every other feature direction stays free to carry clean, discriminative signal. Put briefly, the aim is to confine the damage an attack does to a few directions and give up only those, instead of shrinking everything.
Core Claim
Let $\Phi(X) \in \mathbb{R}^{B \times d}$ be the penultimate-layer features of a batch of $B$ clean images. Let $\Phi(X')$ be the features of their adversarial counterparts, found by the usual inner attack. Define the displacement matrix $D = \Phi(X') - \Phi(X)$. We train with
$$\min_{\theta}\; \mathbb{E}\big[\mathcal{L}_{CE}(f(X),Y) + \beta\,\mathrm{KL}(f(X)\,\|\,f(X'))\big] + \lambda \sum_{i>k} \sigma_i(D),$$
where $\theta$ are the network weights, $\mathcal{L}_{CE}$ is cross-entropy, $\beta$ is the TRADES weight, $\sigma_i(D)$ are the singular values of $D$ in decreasing order, $\lambda$ is the penalty strength, and $k$ is a small number of "sacrificed" directions left unpenalized. Setting $k=0$ recovers the plain nuclear norm. The truncated form with $k>0$ is a weighted nuclear norm with zero weight on the head of the spectrum. It exists to avoid the shrinkage bias that the file identifies as the main failure of plain nuclear-norm penalties.
The claim is falsifiable. At matched AutoAttack robust accuracy on CIFAR-10 and CIFAR-100, this objective should reach higher clean accuracy than TRADES tuned over $\beta$. It should also beat the same objective with the rank penalty replaced by an energy (squared Frobenius) penalty on $D$. The second comparison is the decisive one. If rank and energy penalties give the same trade-off curve, the idea is wrong.
The SVD is of a $128 \times 512$ matrix per step. Its cost is negligible next to the ten-step inner attack, so the method needs none of the SVD-avoidance machinery the nuclear-norm literature has developed.
Why This Is Not Already Done
The work closest to this proposal applies nuclear-norm or spectral penalties to other objects. Domain Generalization via Nuclear Norm Regularization penalizes the nuclear norm of the clean feature matrix $\Phi(X)$. This suppresses weakly label-correlated environmental features and gains 1.7% on DomainBed over standard training. That paper motivates ours by showing that feature-rank penalties select which directions survive. We differ on one point. Penalizing the rank of $\Phi(X)$ removes feature directions for clean inputs too, which is exactly the clean-accuracy cost we want to avoid. Penalizing the rank of $D$ leaves the clean feature space unconstrained and limits only how attacks move it.
Two papers regularize weight matrices, and together they constrain the design. On the Effect of Low-Rank Weights on Adversarial Robustness of Neural Networks adds weight nuclear norms without adversarial training. On small MNIST and Fashion-MNIST networks it gains some robustness, but it falls short of adversarial training and costs clean accuracy (95.08% to 93.27%). Dynamical Low-Rank Compression of Neural Networks with Robustness under Adversarial Attacks argues the opposite direction for weights. It flattens the singular spectrum to control condition number, because naive low rank hurts robustness. We read both as evidence that lowering the rank of parameters is the wrong target. Low weight rank caps capacity for clean and adversarial inputs alike.
Nuclear Norm Regularization for Deep Learning regularizes the input-output Jacobian's nuclear norm cheaply, but it evaluates only on denoising and autoencoders. For a 10-class classifier, the logit Jacobian has rank at most 10 anyway, so its nuclear norm is within a factor of about $\sqrt{10}$ of its Frobenius norm. That makes Jacobian nuclear-norm penalties at the output nearly indistinguishable from ordinary gradient-norm penalties. This is one reason we act on intermediate feature displacements, where the ambient dimension (512) is much larger than the number of classes and rank is a real degree of freedom.
Finally, Theoretically Principled Trade-off between Robustness and Accuracy supplies the base objective, and we keep its inner maximization unchanged. Our proposal does not claim to escape its theoretical barrier on distributions where robust and accurate classifiers are incompatible. It targets the slack on natural image data, where the loss comes from how invariance is achieved rather than from whether it is achievable.
Research Program
The work has three parts, run in order so that the first can kill the idea cheaply.
The first part is a diagnostic on existing models. We train standard TRADES and PGD adversarial training models on CIFAR-10. We compute the effective rank of $D$ (nuclear norm over spectral norm) on held-out batches under AutoAttack perturbations. We also measure how much between-class feature variance lies in the span of $D$'s top singular directions versus its complement.
The hypothesis predicts two things. Displacement in TRADES models is spread across many directions. Clean between-class variance has been lost from those same directions compared with a naturally trained model. If $D$ is already low-rank in TRADES models, the penalty has nothing to do and we stop. On the Effect of Low-Rank Weights on Adversarial Robustness of Neural Networks found that adversarial training lowers weight effective rank, which makes this outcome a genuine possibility.
The second part is the training method above. We sweep $k \in \{0, 4, 16\}$ and $\lambda$ over a small grid, and we apply the penalty at the penultimate layer and optionally at the last residual stage. For each configuration we sweep $\beta$ to trace a full clean-versus-robust curve rather than reporting single points.
The third part is an analysis of where the robustness lives. We measure the head's sensitivity along $D$'s top-$k$ directions against its complement. We check whether an adaptive attack can push displacement out of the learned subspace. We also test on the structurally different trace-norm threat model of Trace-Norm Adversarial Examples, whose FWnucl attack reduced a Madry-trained ResNet-18 from 49.95% (PGD) to 3.06% robust accuracy. This secondary check asks whether confining displacement to a subspace transfers across threat models or overfits to $\ell_\infty$.
Baselines
The naive reference point is standard training without any defense, which gives the clean-accuracy ceiling. Alongside it we run PGD adversarial training at $\ell_\infty$, $\epsilon = 8/255$, as the unmodified robust reference.
The strongest direct competitor is TRADES itself, from Theoretically Principled Trade-off between Robustness and Accuracy. Our method is TRADES plus one term, so any gain must show up against TRADES tuned as carefully as ours. The file reports 88.64% natural accuracy at 56.61% white-box PGD robust accuracy for TRADES on CIFAR-10 with ResNet-18. Those are PGD numbers. No reported AutoAttack result for TRADES was found on Lacuna, so we will rerun it ourselves under AutoAttack. We give TRADES the same number of $\beta$ values, epochs, seeds, data, augmentation and inner-attack steps as our method. This is a total training budget equal to our $(\beta, \lambda, k)$ grid, spent entirely on finer $\beta$ and learning-rate tuning for TRADES.
A second strong competitor is Mitigating Accuracy-Robustness Trade-Off via Balanced Multi-Teacher Adversarial Distillation. It separates clean and robust supervision with two teachers and reports a 1.44% weighted-robust-accuracy gain over TRADES-style training with a ResNet-18 student. The file gives no absolute clean or AutoAttack numbers. We run it with its published balancing scheme at matched student architecture, and we count teacher pre-training in its compute. Because the two methods are compatible, we also report our penalty applied on top of it.
The key ablations isolate the ingredient. The first replaces $\sum_{i>k}\sigma_i(D)$ with $\|D\|_F^2$, an energy penalty on feature displacement, at a $\lambda$ tuned with the same budget. The second adds the clean-feature nuclear norm $\|\Phi(X)\|_*$ from Domain Generalization via Nuclear Norm Regularization to TRADES. The third applies weight nuclear norms in the manner of On the Effect of Low-Rank Weights on Adversarial Robustness of Neural Networks.
Beating a baseline means one of two things, established across three seeds with non-overlapping standard errors. Either the method gains at least one point of clean accuracy at equal AutoAttack robust accuracy, read off interpolated Pareto curves, or it gains at least one point of AutoAttack robust accuracy at equal clean accuracy.
Why not just…?
Why not just tune $\beta$ in TRADES? Tuning $\beta$ moves a model along one trade-off curve. Our claim is that the curve itself moves outward. The fair test is a curve-versus-curve comparison, with TRADES given the full tuning budget described above. If the tuned TRADES curve matches ours, the proposal fails, and the experiment is designed to show that plainly.
Why not just penalize the energy of the feature displacement instead of its rank? This is the most serious threat. A squared Frobenius penalty on $D$ is simpler, differentiable everywhere, and close to existing feature-matching variants of adversarial training. The file's evidence that rank and energy penalties behave differently comes from domain generalization and weight compression (Domain Generalization via Nuclear Norm Regularization, Dynamical Low-Rank Compression of Neural Networks with Robustness under Adversarial Attacks), not from adversarial feature displacement. That evidence is suggestive rather than decisive. The Frobenius-$D$ ablation, run with a matched budget, settles the question. We also expect a mechanistic difference that can be checked directly. Under the energy penalty, clean between-class variance should shrink roughly uniformly across directions. Under the truncated rank penalty, it should shrink only in the $k$ sacrificed directions.
Why not just add the clean-feature nuclear norm from domain generalization to TRADES? That penalty lowers the rank of the clean representation. The most likely effect is more clean accuracy loss, because it removes weakly expressed but real features, which is the failure On the Effect of Low-Rank Weights on Adversarial Robustness of Neural Networks observed for weight penalties. It remains a cheap and plausible competitor, however, because it might discard exactly the brittle features attacks exploit. It is therefore a full ablation arm rather than a strawman.
Why not just use a test-time frequency filter? Contrastive Spectral Rectification: Test-Time Defense towards Zero-shot Adversarial Robustness of CLIP loses only 0.7% clean accuracy on CLIP. It relies on adversarial noise being high-frequency, and the file notes it breaks under low-frequency changes. The low-rank blurs of Trace-Norm Adversarial Examples are such changes. Our FWnucl evaluation reports this contrast directly. We do not claim superiority on the $\ell_\infty$ CLIP setting.
Proposed Experiments and Decisive Checks
The main experiments use CIFAR-10 and CIFAR-100 with ResNet-18 and WideResNet-34-10, at $\ell_\infty$ $\epsilon = 8/255$, with ten-step PGD as the inner attack. Tiny-ImageNet, used by the multi-teacher distillation work, serves as a scale check. All robust accuracy is measured with the full AutoAttack ensemble, following RobustBench: a standardized adversarial robustness benchmark. That work documents robustness overestimates exceeding 10 points in published defenses, so PGD-only numbers will not be used to support any claim.
The first decisive check is the Pareto comparison against tuned TRADES and against the Frobenius-$D$ ablation, as specified in Baselines.
The second decisive check is an adaptive attack built against our defense. It maximizes cross-entropy plus a bonus for displacement orthogonal to the model's top-$k$ displacement subspace, estimated per batch. It also includes a variant that targets the penultimate features directly. If robust accuracy under this attack drops more than two points below AutoAttack, the apparent gain is a gradient-masking or subspace-evasion artifact, and we will report it as such.
The third check is the mechanistic prediction. Relative to TRADES, the method should retain more clean between-class variance in the complement of $D$'s top directions, and it should push the effective rank of $D$ toward $k$. If clean accuracy improves without these signatures, the gain has some other cause and the proposed explanation is wrong.
The fourth check is secondary: FWnucl robust accuracy against the 3.06% Madry reference. It asks whether low-rank displacement generalizes to a threat model the model never trained on.
The main result is positive if the method beats tuned TRADES and the Frobenius-$D$ ablation on the first check, holds up under the second, and shows the signatures of the third.
References
- Domain Generalization via Nuclear Norm Regularization
- On the Effect of Low-Rank Weights on Adversarial Robustness of Neural Networks
- Dynamical Low-Rank Compression of Neural Networks with Robustness under Adversarial Attacks
- Nuclear Norm Regularization for Deep Learning
- Theoretically Principled Trade-off between Robustness and Accuracy
- Trace-Norm Adversarial Examples
- Mitigating Accuracy-Robustness Trade-Off via Balanced Multi-Teacher Adversarial Distillation
- Contrastive Spectral Rectification: Test-Time Defense towards Zero-shot Adversarial Robustness of CLIP
- RobustBench: a standardized adversarial robustness benchmark