Adversarially robust image classifiers are trained by attacking themselves. During training, each image is perturbed in the direction that most increases the loss, and the model learns from the perturbed copy. The largest recent gains in this kind of training did not come from better attacks. They came from more data. Synthetic images from a diffusion model (a generator that learns to turn random noise into realistic samples by removing the noise step by step) raised CIFAR-10 robust accuracy to 64.20% against AutoAttack, and those images beat 80 million real web images (Fixing Data Augmentation to Improve Adversarial Robustness). That synthetic data is produced once, offline, and knows nothing about the classifier. The attacker, by contrast, is adaptive: it queries the deployed model and searches for whatever region the model handles worst. So the training data covers the typical regions of the image distribution, while the attacks that matter land near the decision boundary.
We propose to close that gap by making the data generator act like the attacker. During adversarial training, the diffusion sampler is periodically steered toward realistic, correctly labeled images on which the current model is still fragile under attack, and the new images are added to the training pool. The claim we test is narrow. At a fixed budget of synthetic images and compute, data matched to the model's current failures should buy more robustness to adaptive attack than uniform synthetic data does. The question that decides whether the idea is worth anything is whether steering beats the cheap alternative of generating uniformly and keeping only the hardest samples.
Core Claim
The diffusion-data result is a sample-efficiency story: adversarial training is starved of data, and denser coverage helps (Fixing Data Augmentation to Improve Adversarial Robustness). If that is right, the most valuable extra samples are those in the regions an adaptive adversary exploits, and uniform sampling spends most of its budget elsewhere. We expect steered generation to reach a given AutoAttack robust accuracy with several times fewer synthetic images than uniform generation. We also expect it to beat hard-example selection from a uniform pool, specifically in low-density regions that the uniform pool rarely contains. If selection matches steering, we will report that the benefit comes from matching the data to failures, not from generation itself. That is still a useful and cheaper result.
How This Differs From Prior Work
Two lines of work motivate the proposal. The first makes augmentation adaptive, but only inside a fixed menu of low-level operations. AROID learns a per-image policy over rotations, crops, and color changes. Its policy is rewarded for raising the model's loss under attack, it uses an auxiliary "affinity" model to keep images natural, and it shrinks robust overfitting from 5.64% to 0.91%. Universal Adaptive Data Augmentation picks loss-maximizing augmentation parameters online through simulated gradients, but it never measures adversarial robustness. Compositional Adversarial Training learns sequences of image operations against watermark decoders. It breaks down beyond two steps because the images become unrecognizable. All three search over pixel-level edits of existing images. We search over new images drawn from a generative model, which can change pose, shape, and context while staying realistic.
The second line supplies the data source and the stabilizer. Data Augmentation Can Improve Robustness shows that weight averaging is essential: it keeps a running average of model weights, which acts as an ensemble over training time. The same paper shows that CutMix with weight averaging reaches 60.07% on CIFAR-10. The diffusion follow-up then decouples generation from training entirely. We keep its weight averaging and its diffusion model and change only how samples are chosen.
Several other papers constrain the design. The "spectral chameleon" finding in A Spectral Perspective towards Understanding and Improving Adversarial Robustness is that white-box attacks move into whichever frequency bands the model leaves undefended. This suggests that failure regions shift as training proceeds, so steering must be refreshed rather than done once. The trade-off in A Fourier Perspective on Model Robustness in Computer Vision warns that gains against one perturbation type can cost up to 30% on low-frequency corruptions such as fog and contrast changes. We therefore must check corruption robustness, not only adversarial accuracy. Finally, the many defenses that collapse under attacks designed to bypass them (DeepCloak, and the expectation-over-transformation results for DBD) mean any reported gain must survive strong adaptive evaluation and explicit checks for gradient masking. Gradient masking is when a model hides useful gradients, so attacks fail without the model being truly robust.
Research Program
We start from a class-conditional diffusion model trained on CIFAR-10, the same kind used to produce the 1M-image pool. Every $K$ epochs of adversarial training, we draw a fresh batch of samples. In the final denoising steps, each sample is guided by the gradient of a steering score evaluated on the sampler's current estimate $\hat{x}_0$ of the clean image:
$$ s(\hat{x}_0, y) = \max_{\|\delta\|_\infty \le \epsilon} \mathrm{CE}\big(f_\theta(\hat{x}_0+\delta), y\big) \;-\; \mu\, \mathrm{CE}\big(g(\hat{x}_0), y\big) $$
The symbols are defined as follows. $f_\theta$ is the classifier being trained, read from its weight-averaged copy. $y$ is the conditioning class. The inner maximum is approximated with a few steps of PGD, the projected-gradient attack. $\mathrm{CE}$ is cross-entropy. $g$ is a frozen, independently trained label-check classifier, and $\mu$ weights how strongly the image must still look like class $y$ to that classifier. The first term pulls samples toward images the current model fails on under attack. The second term plays the role of AROID's affinity model: it stops the sampler from producing ambiguous images whose label is wrong, which would merely inject label noise. Steered samples replace a fixed fraction of the synthetic pool at each refresh, and training mixes real, uniform synthetic, and steered images.
Three analyses make the mechanism visible. The first is a coverage analysis. We embed steered and uniform samples with an independent feature extractor and measure how much of the steered set falls in regions where uniform samples are sparse. This tells us whether steering reaches places selection cannot. The second is a label-fidelity audit. Human raters review a random subset of steered images and judge whether the class is preserved, and we track how the rate of ambiguous images changes with $\mu$. The third is a spectral diagnostic. Following the Fourier-sensitivity method of Yin et al., we test whether steered training moves the model's vulnerability across frequency bands, and whether the chameleon behavior of attacks weakens.
Compute is the central fairness issue. Guided sampling backpropagates through the classifier at each guided step, so a steered image costs more than a uniform one. All comparisons are therefore reported at matched total GPU-hours, generation and training combined. The uniform arm receives its extra budget as additional samples.
Baselines
The naive reference point is adversarial training on real CIFAR-10 with standard flips and crops and with weight averaging. It uses the same WideResNet-28-10, the same perturbation radius $\epsilon = 8/255$, and the same number of training epochs as every other arm. This tells us how much any data intervention adds.
The strongest existing method is uniform DDPM synthetic data with weight averaging, the recipe from Fixing Data Augmentation to Improve Adversarial Robustness. That paper reports 64.20% AutoAttack robust accuracy on CIFAR-10 at $L_\infty$, $\epsilon=8/255$, with 1M synthetic images. It is the right competitor because our proposal changes only how synthetic samples are chosen. We will rerun it ourselves with the same diffusion checkpoint, architecture, schedule, and hyperparameter-search budget. We will do so at pool sizes of 50k, 200k, and 1M, which gives a sample-efficiency curve rather than a single point. Beating it means one of two outcomes, each holding over at least three seeds. Either steered training matches the uniform 1M result with a pool several times smaller at equal or lower total compute, or it exceeds uniform data by at least one AutoAttack point at matched compute.
Two secondary references come from the file. The first is CutMix with weight averaging from Data Augmentation Can Improve Robustness, reported at 60.07% on CIFAR-10 and 32.43% on CIFAR-100. The second is AROID, reported at 55.91% on CIFAR-10. We run AROID through its official code, both alone and stacked on the uniform synthetic pool, so that adaptive pixel-level augmentation is tested on the same data. For CIFAR-100 we found no reported diffusion-data result on Lacuna, so we will establish that uniform-DDPM baseline ourselves.
The key ablation is hard-example selection. We generate a large uniform pool, and at each refresh we keep the samples with the highest adversarial loss under the current model, at the same refresh rate and compute. This isolates the generation step from the matching step. A second ablation sets $\mu = 0$ to measure how much the label check matters. A third freezes the steering classifier at an early checkpoint, which tests whether refreshing matters or whether one round of steering suffices.
Why not just…?
Why not just generate more uniform samples? This is the default answer, and scale clearly helps, given the 64.20% result. Our sample-efficiency curve answers it directly. If uniform data at matched GPU-hours closes the gap at every pool size, steering is not worth its cost, and we will say so. We expect steering to win most clearly at small and medium pool sizes, where uniform sampling has the fewest boundary-region images.
Why not just keep the hardest images from a uniform pool? This is the most serious threat, because it is cheap and targets the same failure regions. It can only reweight what the uniform pool already contains, and the coverage analysis measures how much of steering's output lies outside that pool's dense regions. If selection matches steering, the honest conclusion is that matching data to failures is what helps. We would then recommend selection over guided generation.
Why not just use adaptive augmentation like AROID on the synthetic data? AROID's search is limited to low-level operations applied to existing images. CAT's experience suggests that pushing such pipelines harder destroys the image content rather than finding new realistic hard cases. Running AROID stacked on uniform synthetic data, as described above, settles whether pixel-level adaptivity already captures the gain.
Why not just raise the perturbation radius or the attack strength during training? A larger $L_p$ ball (a fixed-size limit on how much pixels may change) still explores only near existing images. Steering changes the image itself along the data distribution. If stronger attacks during training gave the same AutoAttack gain without the corruption-robustness cost, steering would be redundant. We include an arm with a larger training radius and a matched budget to check this.
Decisive Experiments
All models are evaluated on the RobustBench protocol with full AutoAttack. Adaptive Auto Attack ($A^3$) serves as a second, independently tuned adaptive attacker. The IndicatorsOfAttackFailure checks rule out gradient masking. Steered training should not mask gradients, since the deployed model is an ordinary network, but because the gain is measured by attack failure we will verify this rather than assume it.
Three results decide the proposal. The first is the matched-compute sample-efficiency curve against uniform DDPM. The second is the head-to-head against hard-example selection. The third is the corruption check on CIFAR-10-C, where Yin et al. report a mean corruption error of 100 for natural training and 64 for AutoAugment. A steered model that gains AutoAttack accuracy while losing ground on fog and contrast would reproduce the Fourier trade-off rather than escape it. We also track the gap between peak and final robust accuracy, using AROID's 0.91% as the reference for successful control of robust overfitting.
If steering wins on the first two and holds on the third, the evidence is that adversarial training should draw its data from the regions an adaptive attacker would target. If selection ties, the field gets a cheaper recipe and a clear negative result on guided generation.
References
- Fixing Data Augmentation to Improve Adversarial Robustness
- AROID: Improving Adversarial Robustness Through Online Instance-Wise Data Augmentation
- Universal Adaptive Data Augmentation
- Compositional Adversarial Training for Robust Visual Watermarking
- Data Augmentation Can Improve Robustness
- A Spectral Perspective towards Understanding and Improving Adversarial Robustness
- A Fourier Perspective on Model Robustness in Computer Vision
- DeepCloak: Masking Deep Neural Network Models for Robustness Against Adversarial Samples
- Adversarial Attacks Already Tell the Answer: Directional Bias-Guided Test-time Defense for Vision-Language Models
- RobustBench: a standardized adversarial robustness benchmark
- Practical Evaluation of Adversarial Robustness via Adaptive Auto Attack