Adversarial training makes a classifier robust by training it on inputs that have been deliberately perturbed to fool it. The best perturbations come from a multi-step attack, projected gradient descent (PGD), which costs roughly ten times a normal training step. The cheap substitute is a single gradient step (FGSM), but it has a known failure. Partway through training, the model reshapes its own loss surface so that one-step attacks stop finding anything. The model then keeps training on these weakened examples, and its accuracy against real attacks collapses to near zero. This failure is called catastrophic overfitting (Understanding Catastrophic Overfitting in Single-step Adversarial Training). We argue it is a case of semantic drift, meaning the slow decay that sets in whenever a system trains on signals it generates itself. The cheap attack is a pseudo-label for the expensive one. As the model changes, that pseudo-label becomes less faithful, and training on it makes the mismatch worse.
Semi-supervised learning already has a precise tool for this situation. Doubly-Robust Self-Training trains a student mostly on a teacher's cheap pseudo-labels. It then uses a small set of true labels to subtract out the teacher's bias. If the teacher becomes useless, training falls back to the labeled data alone instead of collapsing. We propose the same construction for adversarial training. Every example gets a cheap one-step attack. A small random subset also gets an expensive PGD attack. The PGD subset is used to cancel the bias of the one-step loss, so the combined objective is an unbiased estimate of the full PGD loss at a fraction of its cost. The gap between the two attacks on the subset doubles as a live gauge of drift, and it can tell training when to buy more expensive attacks.
Core Claim
Fix a minibatch $B$. For each example $x$ in it, let $g(x)$ be the training loss on the one-step perturbation of $x$ and $h(x)$ the loss on its PGD perturbation. Draw a uniform random subset $S \subset B$ of size $m$ and train on
$$ \hat{L}_\lambda = \frac{1}{m}\sum_{x\in S} h(x) + \lambda\Big(\frac{1}{|B|}\sum_{x\in B} g(x) - \frac{1}{m}\sum_{x\in S} g(x)\Big), $$
where $\lambda$ weights the cheap-attack correction.
For any $\lambda$, the expectation of $\hat{L}_\lambda$ over the random subset equals the full-batch PGD loss $\frac{1}{|B|}\sum_{x\in B} h(x)$. This holds because the bracketed term has zero mean when $S$ is drawn uniformly. Setting $\lambda=1$ recovers the form used by Zhu et al. Setting $\lambda=0$ is plain PGD training on the subset. The variance of $\hat{L}_\lambda$ is smallest when $\lambda$ equals the regression slope of $h$ on $g$ within the batch, which is the standard control-variate result.
This gives the claim a specific shape. While the one-step attack still tracks PGD ($g$ and $h$ strongly correlated), the estimator gets near-PGD-quality gradients for roughly the price of the one-step attack. When catastrophic overfitting begins, $g$ decouples from $h$. The fitted $\lambda$ then shrinks toward zero, and the objective reverts to honest PGD training on $S$ instead of following the broken proxy. The proxy can therefore stop helping, but it cannot steer training wrong.
The compute cost per step is about $2 + \tfrac{m}{|B|}(K+1)$ forward-backward passes for a $K$-step PGD attack. For $m/|B| = 1/8$ and $K = 10$ that is about 3.4 passes, against 11 for full PGD-10.
Positioning
Three lines of work motivate the drift reading of catastrophic overfitting.
- Ding et al. show that misaligned self-generated training signals accelerate decline instead of staying isolated.
- Zverev et al. prove that training on recursive self-generations produces systematic bias plus exploding variance.
- Together these say that fixing the bias alone is not enough. Our estimator targets bias exactly through the unbiased construction, and controls variance through $\lambda$ and the subset size $m$.
Doubly-Robust Self-Training supplies the method. We take its estimator and make three changes:
- We move it from labels to perturbations. Its "labeled set" becomes a freshly sampled PGD subset in every batch, not a fixed annotated pool.
- We let $\lambda$ adapt.
- We add a budget rule driven by the measured drift. Zhu et al. report that their estimator needs curriculum schedules for stability in deep networks. This is a real constraint, because the negatively weighted subset term injects gradients that push toward higher one-step loss on $S$. We test for exactly this instability.
Two other routes to cheap robustness bypass the question rather than answering it.
- Sustainable Self-evolution Adversarial Training limits expensive retraining to a replay buffer chosen by Monte Carlo uncertainty. That buffer stores raw adversarial images and is costly to select, whereas our subset is random and stateless.
- Purifiers such as OSCP and AMRM-Pure avoid training cost by moving the work to inference time.
- Parameter-efficient defenses such as Robustness Tokens shrink the set of trained parameters but still need an inner attack loop.
Our estimator is complementary to the last of these. It reduces the cost of that inner loop, so we also test it there.
Research Program
The program has five parts.
Implement the estimator. We build it on top of FGSM with random initialization in two variants. The fixed variant uses $\lambda=1$. The adaptive variant re-estimates $\lambda$ every step from the $(g,h)$ pairs on $S$, clipped to $[0,1]$ and smoothed over steps.
Instrument the drift gauge. At each step we log the subset statistic $\frac{1}{m}\sum_{x \in S}(h(x)-g(x))$ and the within-batch correlation of $g$ with $h$. We then ask whether these signals rise before PGD validation accuracy collapses. Catastrophic overfitting is usually caught today only after the collapse, so an early warning is a contribution even independent of the estimator.
Make the budget adaptive. We increase $m$ when the gap exceeds a threshold and decrease it when the two losses agree. This spends PGD compute only when the cheap proxy has drifted.
Test whether the bias correction is what helps. We compare against a mixed batch at identical $m$. This compares the estimator against simply training on the PGD subset alongside one-step examples.
Port to the large-model setting. We use the estimator in the inner loop of Robustness Tokens training on a frozen DINOv2 backbone, where each PGD step through a giant transformer is the dominant cost.
Baselines
Naive reference points. These are plain FGSM with random initialization, the standard fast method described in the catastrophic overfitting literature (Understanding Catastrophic Overfitting in Single-step Adversarial Training), and standard training with no adversarial examples. They bound the result from below and show where collapse happens.
Strongest robustness reference. This is full PGD-10 adversarial training under the identical schedule, architecture, and augmentation. It is the quality ceiling that the estimator is trying to approximate. We found no reported result on Lacuna for PGD-10 or FGSM adversarial training robust accuracy on CIFAR-10 under AutoAttack. All of these numbers will be produced in-house with one shared codebase. For context only, the strongest CIFAR-10 numbers in the file come from different protocols: AMRM-Pure reports 75.83% robust accuracy under $\ell_\infty$ attacks for its robust MaskDiT purifier, and SSEAT reports 74.86% accuracy against FGSM. Neither is directly comparable to an AutoAttack evaluation of a trained classifier, and we will say so.
Strongest cost-matched competitor. This is PGD with as many steps as fit the estimator's compute, roughly PGD-2 or PGD-3 applied to the whole batch. It answers the obvious objection that one could just run a shorter attack on everything. Fair comparison means equal wall-clock time and equal forward-backward passes per epoch, with the same tuning budget for step size and learning rate.
Key ablation. This is the mixed batch: one-step loss on $B \setminus S$ plus PGD loss on $S$, with no subtraction term. It is biased toward the one-step objective, and it isolates the bias correction as our key ingredient. A second ablation sets $\lambda=1$ with no adaptation, isolating the adaptive weight.
Large-model testbed. The baseline is Robustness Tokens as published. Pulfer et al. report 63.1% ImageNet white-box accuracy for DINOv2-G (0.3% without defense) and 85.2% clean accuracy, in 200 steps. Beating it there means matching that robust accuracy within about one point at substantially fewer backbone forward-backward passes.
On CIFAR, beating the strongest competitor means one of two outcomes. Either the method beats cost-matched PGD-2/3 on AutoAttack accuracy at equal compute, or it comes within one to two points of PGD-10 at about a third of its cost. In addition, it must show zero catastrophic-overfitting runs across seeds in settings where FGSM and the mixed batch collapse.
Why not just…?
Why not just mix in a fraction of PGD examples? This is the most serious threat. A small PGD share might be enough to keep the loss surface from going degenerate, which would make the subtraction term unnecessary. The mixed-batch ablation at identical $m$ settles it directly. If the mixed batch matches the estimator on robust accuracy and collapse rate at every perturbation radius, the bias-correction claim fails. What would survive is the drift gauge and the adaptive budget, and we would reframe the work around them. Our expectation favors the estimator at large radii. The mixed objective is still dominated by the one-step loss on most of the batch, and that is exactly the part that becomes uninformative during collapse.
Why not just run a shorter PGD attack on every example? Two- or three-step PGD is a strong, simple, cost-matched option. It is a main baseline here, not a strawman. Its weakness is that it is a biased approximation whose bias is never measured. Our estimator is unbiased for the ten-step loss and reports its own error through the gap statistic. The equal-compute comparison will show whether unbiasedness is worth anything in practice.
Why not just detect overfitting and restart, or switch to PGD when it happens? Early stopping on a PGD validation set works, but it discards the remaining training budget and only notices collapse after it happens. The adaptive-$m$ variant is a smooth, per-step version of switching to PGD. We will compare against a hard switch triggered by the same gap statistic, so any advantage has to come from the smoothness rather than from the signal.
Why not just avoid training cost entirely with prompts or purification? These are live alternatives with weaknesses already documented in the file.
- Bhagwatkar et al. raise LLaVA-7B accuracy under an adaptive PGD attack from 12.54% to 91.46% with a single defensive sentence, but they note it breaks once the attacker includes that prompt in the optimization.
- Purifiers trade training cost for inference cost and remain open to adaptive attacks (AMRM-Pure).
Our question is narrower: how to make robust training itself cheaper. These methods sit outside that comparison.
Proposed Experiments
Datasets and models. We use CIFAR-10 and CIFAR-100 with a PreAct ResNet-18 and a WideResNet. Attacks are $\ell_\infty$ with radii from 8/255 to 16/255. The upper end of that range is where one-step training is known to collapse frequently.
Evaluation. We report clean accuracy, PGD-50 accuracy, and AutoAttack accuracy. We also report wall-clock time and forward-backward passes per epoch, and the fraction of seeds (out of at least five) that show catastrophic overfitting.
Drift diagnostic. For each run we record the gap and correlation traces. We measure how many epochs of warning the gap gives before robust validation accuracy drops below half its peak.
Large-model test. We repeat the Robustness Tokens setup on ImageNet with DINOv2 and train the tokens using the estimator in place of full PGD. We also report the segmentation transfer on ADE20K that Pulfer et al. measure (mIoU under attack: 4.7 without defense, 23.3 with tokens).
Decisive checks. There are three.
- At 16/255 on CIFAR-10, the estimator must avoid collapse on every seed where both FGSM and the equal-$m$ mixed batch collapse on most seeds.
- At equal compute, it must exceed PGD-2/3 on AutoAttack accuracy.
- The gap statistic must rise measurably before the robust-accuracy collapse in the FGSM runs.
Failure criteria. Any of the following would count against the method: the mixed batch matching it, the fixed $\lambda=1$ variant being unstable without a curriculum (as Zhu et al. warn), or the gap failing to lead collapse. Each of these is a clean negative result about whether reading catastrophic overfitting as self-induced drift is useful.
References
- Understanding Catastrophic Overfitting in Single-step Adversarial Training
- Doubly-Robust Self-Training
- Combating confirmation bias: a unified pseudo-labeling framework for entity alignment
- On the Dangers of Bootstrapping Generation for Continual Learning and Beyond
- Sustainable Self-evolution Adversarial Training
- Instant Adversarial Purification with Adversarial Consistency Distillation
- AMRM-Pure: Semantic-Preserving Adversarial Purification
- Robustness Tokens: Towards Adversarial Robustness of Transformers
- Improving Adversarial Robustness in Vision-Language Models with Architecture and Prompt Design