TL;DR
This paper introduces a fine-grained approach to understanding why Convolutional Neural Networks (CNNs) are vulnerable to adversarial attacks by applying Shapley values to the frequency domain. The authors discover that robustness isn't just a simple matter of "low vs. high" frequencies. Instead, it depends on how specific frequency components contribute positively or negatively to a prediction. This insight leads to a new data augmentation technique—Class-wise Shapley value-guided Augmentation (CSA)—that significantly boosts model defense.
What Problem Does This Paper Address?
Deep learning models, specifically CNNs, are notoriously "brittle." Small, invisible perturbations (adversarial noise) can easily trick them into misclassifying images. A popular theory explaining this vulnerability is frequency bias: while humans rely on shapes and objects (low frequencies) to recognize images, CNNs often "cheat" by relying heavily on high-frequency textures (HFCs). These high-frequency components are invisible to humans but easily corrupted by attackers.
However, previous research usually just chopped the image spectrum into two rough buckets: "low" and "high." This paper argues that this binary approach is too simplistic. It fails to account for instance disparity (the fact that different images rely on different frequencies) and doesn't explain why some attacks succeed even when they only target low frequencies. The authors seek a mathematically rigorous way to quantify exactly which frequencies help or hurt a model's decision-making process.
Methods
The researchers bridge the gap between computer vision and cooperative game theory to quantify frequency impact:
- Frequency Domain Mapping: They use the Discrete Fourier Transform (DFT) to move images from the spatial domain (pixels) to the frequency domain.
- Shapley Values as a Metric: In game theory, the Shapley value determines how to fairly distribute a "payout" among players in a coalition. Here, the "players" are individual frequency components, and the "payout" is the model's confidence in the correct class.
- Quantification: Because calculating exact Shapley values is computationally expensive, they use Monte Carlo Permutation Sampling to estimate the contribution of each frequency.
- Class-wise Shapley value-guided Augmentation (CSA): Based on their analysis, the authors identify Negative Frequency Components (NFCs)—frequencies that actively reduce the model's confidence. They propose a training strategy where they take these "confusing" features from one image and add them to another image of the same class. This forces the model to learn to ignore these misleading signals during training.
Key Insights and Ideas
- Instance Disparity: The reliance on specific frequencies is not a one-size-fits-all rule. For example, a model might rely on low-frequency color blobs to classify a "goldfish," but rely on high-frequency fin contours for a "shark." The "frequency bias" varies significantly from image to image.
- The Cost of Robustness: Adversarial Training (AT) makes models robust by essentially "sacrificing" HFCs. AT models learn to ignore high frequencies almost entirely to avoid being tricked. This explains why robust models often suffer a drop in "clean" accuracy compared to standard models.
- The Fairness Gap: The authors found a negative correlation between how much a class relies on HFCs and how robust it becomes after training. "Transportation" classes (like cars and planes) are often more robust than "Animal" classes because the latter rely more heavily on vulnerable high-frequency textures.
- Why LFC Attacks Work: Even though HFCs are generally more vulnerable, attacks on low-frequency components (LFCs) can still succeed. They work by "boosting" the impact of existing Negative Frequency Components that were already present in the clean image.
Key Results
The proposed CSA augmentation consistently improved robustness across different architectures (ResNet-18, VGG16, WideResNet) evaluated on standard benchmarks like CIFAR-10, CIFAR-100, and Tiny ImageNet:
- Robustness Gains: On CIFAR-10, adding CSA to standard Adversarial Training (PGD-AT) improved robust accuracy against PGD-20 attacks from 46.38% to 49.42%.
- Efficiency: Unlike some explainability methods that require thousands of samples, the authors showed that extracting NFCs from just 3% of the training data was enough to yield significant robustness improvements.
- Model Agnostic: The method worked across various model types, suggesting that these frequency-based insights reflect a fundamental property of how CNNs process visual information.
Limitations and Future Work
- Architecture Scope: The study focused almost exclusively on CNNs. As Vision Transformers (ViTs) become more dominant, it remains to be seen if the same frequency-domain Shapley logic applies to attention-based mechanisms.
- Computational Overhead: While the authors optimized the process using Monte Carlo sampling, calculating Shapley values is still more computationally intensive than standard training procedures.
- Future Directions: The authors suggest extending this framework to other domains, such as Graph Neural Networks (GNNs), and exploring more sophisticated ways to "realign" features beyond simple additive augmentation.
Summary updated