Abstract
Despite the significant advances that deep neural networks (DNNs) have achieved in various visual tasks, they still exhibit vulnerability to adversarial examples, leading to serious security concerns. Recent adversarial training techniques have utilized inverse adversarial attacks to generate high-confidence examples, aiming to align the distributions of adversarial examples with the high-confidence regions of their corresponding classes. However, our investigation reveals that high-confidence outputs under inverse adversarial attacks are correlated with biased feature activation. Specifically, training with inverse adversarial examples causes the model's attention to shift towards background features, introducing a spurious correlation bias. To address this bias, we propose Debiased High-Confidence Adversarial Training (DHAT), a novel approach that not only aligns the logits of adversarial examples with debiased high-confidence logits obtained from inverse adversarial examples, but also restores the model's attention to its normal state by enhancing foreground logit orthogonality. Extensive experiments demonstrate that DHAT achieves state-of-the-art performance and exhibits robust generalization capabilities across various vision datasets.
Introduction: The Spurious Correlation Bias in Adversarial Training
Adversarial training (AT) is the gold standard for defending deep neural networks against malicious perturbations. Recent high-performance AT methods often use "inverse adversarial attacks"—perturbations that increase model confidence—to create high-quality targets for aligning adversarial examples.
However, this research identifies a "concerning phenomenon": these inverse attacks often trick the model into focusing on irrelevant background features. For example, a model might identify a sheep with high confidence not by its shape, but by the grass in the background. This spurious correlation bias leads to "robust overfitting," where the model performs well on training data but fails to generalize because its robustness is built on non-causal background features. Quantitative analysis using Intersection over Union (IoU) confirms that standard AT models are often "distracted" by the background as attack strength increases.
Methodology: Debiased High-Confidence Adversarial Training (DHAT)
To address this bias, the authors propose Debiased High-Confidence Adversarial Training (DHAT), which introduces two novel regularization components to the standard AT pipeline:
- Debiased High-Confidence Logit Regularization (DHLR): This technique uses saliency maps (like Grad-CAM) to identify background regions. It calculates the specific contribution of these background features to the model's output (the "biased logits") and subtracts them from the high-confidence target. This ensures the model aligns adversarial examples with "clean," object-focused targets.
- Foreground Logit Orthogonal Enhancement (FLOE): To further decouple foreground and background influences, FLOE uses a projection-based approach in the affine space. It minimizes the correlation between the model's high-confidence predictions and the background feature representations, effectively forcing the model to rely on meaningful foreground information.
Experimental Results and Generalization
DHAT was evaluated on major computer vision benchmarks including CIFAR-10, CIFAR-100, and ImageNet-1K. The method consistently outperformed existing state-of-the-art (SOTA) methods across multiple architectures like WRN28-10 and ResNet-50.
- Superior Robustness: On CIFAR-10, DHAT achieved a PGD-10 accuracy of 62.67% (when combined with CFA), significantly higher than previous methods like MART or UIAT.
- Reduced Robust Gap: A critical insight of this work is the reduction of the "Robust Gap"—the difference in robustness between training and test sets. On ImageNet-1K, DHAT achieved a gap of only 9.53%, compared to 14.68% for the previous SOTA.
- Stability: DHAT maintains higher accuracy even as the strength of the adversarial attack ($\epsilon$) increases, demonstrating better stability than its competitors under both PGD and C&W attacks.
Conclusion
DHAT addresses a hidden flaw in modern adversarial defense: models often achieve high confidence by latching onto irrelevant background features rather than actual objects. By explicitly debiasing high-confidence targets and enforcing foreground focus, DHAT achieves state-of-the-art robustness and significantly better generalization across benchmarks. Future research could investigate if similar spurious correlation biases exist in other domains like 3D point clouds or natural language processing.
Summary updated