Failure Cases Are Better Learned But Boundary Says Sorry: Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training

Yanyun Wang, Li Liu

Introduction

Adversarial Training (AT) is the primary defense against adversarial attacks, yet it is plagued by a persistent trade-off between clean accuracy and robust performance. Conventional wisdom suggests that "hard" adversarial examples—those that continue to fool a model post-training—are simply under-learned. This paper challenges that assumption, proposing that these failure cases are actually "over-learned" to be too consistent, which inadvertently degrades the decision boundary. To resolve this, the authors introduce Robust Perception Adversarial Training (RPAT).

The "Over-Learning" Paradox

By analyzing the internal responses (logits) of models during training, the authors discovered a surprising trend. In standard PGD-AT, failure cases—samples the model fails to classify correctly under attack—actually exhibit lower Mean Squared Error (MSE) in their logits compared to successfully defended samples.

MSE analysis showing that failure cases in PGD-AT have more consistent perception (lower MSE) than success cases.

This suggests that the model becomes hyper-focused on maintaining "perception consistency," treating adversarial signals as noise to be ignored. This over-consistency prevents the model from utilizing the perturbation's information to correctly position its decision boundary, leading to the "Boundary says sorry" phenomenon where the boundary is squeezed into sub-optimal locations.

Robust Perception Adversarial Training (RPAT)

Instead of forcing the model to be invariant to small perturbations (which causes gradient signals to vanish), RPAT encourages the model's internal perception to change smoothly and proportionally to the input change.

The RPAT framework: using interpolation and a ratio-based MSE loss to ensure smooth perception transitions.

The core of RPAT is a regularization term that ensures the ratio of perception change in logit space matches the ratio of the input perturbation. If an input is interpolated between a benign sample $x$ and an adversarial sample $x'$, the model's output for the interpolated point should ideally fall linearly between the outputs for $x$ and $x'$. This approach maintains a healthy gradient signal, allowing the model to learn a smoother, more effective decision boundary. Theoretically, this objective regularizes the Jacobian and Hessian of the network, promoting Lipschitz continuity.

Empirical Results

RPAT was evaluated on CIFAR-10, CIFAR-100, and Tiny-ImageNet using architectures like ResNet-18 and WideResNet-34-10. Key results include:

  • CIFAR-10 (WideResNet-34-10): Achieved 87.57% clean accuracy and 55.79% Auto-Attack (AA) robustness, setting a new benchmark for the accuracy-robustness trade-off.
  • Generalization: The method proved robust across both $\ell_\infty$ and $\ell_2$ threat models, consistently outperforming established baselines like TRADES and MART.

Conclusion

The findings suggest that total invariance is a sub-optimal goal for adversarial robustness. By facilitating smooth perception changes, RPAT allows models to better navigate the trade-off between accuracy and security, providing a more principled approach to building robust deep learning systems.

Summary updated

BibTeX
@inproceedings{wang2025failure,
  author = {Yanyun Wang and Li Liu},
  title = {{Failure Cases Are Better Learned But Boundary Says Sorry: Facilitating Smooth Perception Change for Accuracy-Robustness Trade-Off in Adversarial Training}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2025},
}
Figure