Abstract
Deep Neural Networks (DNNs) are known for their vulnerability to adversarial attacks. This study challenges the conventional "spectral bias" theory, which posits that networks primarily rely on low-frequency information. By analyzing feature extraction across the Fourier domain, we demonstrate that effective features exist throughout the frequency spectrum. However, a "robustness paradox" exists: while low-frequency features are more stable, high-frequency features dominate the final decision-making process, creating a significant vulnerability.
Introduction
The brittleness of DNNs to imperceptible perturbations remains a central challenge in computer vision. Previous research suggested that networks learn broad shapes (low frequencies) first and struggle with fine textures (high frequencies). This paper investigates how models actually integrate these disparate frequency bands and the resulting implications for adversarial robustness.
Methodology
To probe the spectral characteristics of DNNs, the authors utilize two novel approaches:
- Frequency-Specific Training: Models are retrained on ImageNet data filtered into specific frequency bands. This tests whether networks can achieve high accuracy using only restricted spectral information, such as high-frequency edges or mid-range textures.
- F-PGD-L Attack: A refined Fourier-based Projected Gradient Descent attack. Unlike standard methods that may "leak" information across frequencies due to aliasing, F-PGD-L ensures perturbations are strictly confined to the target frequency band, allowing for precise sensitivity analysis.
Key Insights: The Distribution of Effective Features
The study reveals that "effective features"—those sufficient for correct classification—are not confined to low frequencies. As shown in the frequency sensitivity analysis below, models retrained on specific bands can achieve significant accuracy (over 50% Top-1) even in narrow high-frequency ranges. This suggests that standard models are "lazy," utilizing only the minimum low-frequency information necessary while ignoring a wealth of available high-frequency features.
Furthermore, the authors identify a "High-Frequency Dominance" effect. In scenarios where low-frequency and high-frequency cues conflict, the network's final prediction is almost invariably dictated by the high-frequency components.
Results: The Robustness Paradox
The investigation into adversarial robustness reveals a critical contradiction. While low-frequency features are inherently more stable and resistant to noise, the model's reliance on high-frequency details for its final decision makes the entire system vulnerable.
As illustrated above, targeted attacks on high-frequency bands cause a much sharper decline in model confidence compared to low-frequency attacks. This confirms that high-frequency features are the primary "Achilles' heel" of modern DNN architectures.
Conclusion
This research provides a new perspective on adversarial vulnerability, moving beyond the idea of high-frequency "noise" to recognize high-frequency "features" as the dominant but fragile basis for model decisions. Future work should focus on developing architectures that better integrate information across the frequency spectrum to enhance stability without sacrificing the discriminative power of fine details.
Summary updated