Publication Date
1-1-2024
Journal
Frontiers in Computational Neuroscience
DOI
10.3389/fncom.2024.1387077
PMID
38966128
PMCID
PMC11222393
PubMedCentral® Posted Date
6-20-2024
PubMedCentral® Full Text Version
Post-print
Published Open-Access
yes
Keywords
adversarial examples, implicit regularization, neural networks, convolutional architectures, Uncertainty Principle
Abstract
Adversarial attacks are still a significant challenge for neural networks. Recent efforts have shown that adversarial perturbations typically contain high-frequency features, but the root cause of this phenomenon remains unknown. Inspired by theoretical work on linear convolutional models, we hypothesize that translational symmetry in convolutional operations together with localized kernels implicitly bias the learning of high-frequency features, and that this is one of the main causes of high frequency adversarial examples. To test this hypothesis, we analyzed the impact of different choices of linear and non-linear architectures on the implicit bias of the learned features and adversarial perturbations, in spatial and frequency domains. We find that, independently of the training dataset, convolutional operations have higher frequency adversarial attacks compared to other architectural parameterizations, and that this phenomenon is exacerbated with stronger locality of the kernel (kernel size) end depth of the model. The explanation for the kernel size dependence involves the Fourier Uncertainty Principle: a spatially-limited filter (local kernel in the space domain) cannot also be frequency-limited (local in the frequency domain). Using larger convolution kernel sizes or avoiding convolutions (e.g., by using Vision Transformers or MLP-style architectures) significantly reduces this high-frequency bias. Looking forward, our work strongly suggests that understanding and controlling the implicit bias of architectures will be essential for achieving adversarial robustness.