Adversarial Attacks Leverage Interference Between Features in Superposition
This provides a mechanistic explanation for adversarial vulnerability as a byproduct of representational compression, addressing a fundamental problem in AI safety for researchers and practitioners.
The paper tackles the origin of adversarial examples in neural networks by arguing that vulnerability stems from efficient information encoding through superposition, where networks represent more features than dimensions, and demonstrates that adversarial perturbations exploit interference between superposed features, showing this suffices to create vulnerability in synthetic settings and persists in a ViT on CIFAR-10.
Fundamental questions remain about when and why adversarial examples arise in neural networks, with competing views characterising them either as artifacts of the irregularities in the decision landscape or as products of sensitivity to non-robust input features. In this paper, we instead argue that adversarial vulnerability can stem from efficient information encoding in neural networks. Specifically, we show how superposition - where networks represent more features than they have dimensions - creates arrangements of latent representations that adversaries can exploit. We demonstrate that adversarial perturbations leverage interference between superposed features, making attack patterns predictable from feature arrangements. Our framework provides a mechanistic explanation for two known phenomena: adversarial attack transferability between models with similar training regimes and class-specific vulnerability patterns. In synthetic settings with precisely controlled superposition, we establish that superposition suffices to create adversarial vulnerability. We then demonstrate that these findings persist in a ViT trained on CIFAR-10. These findings reveal adversarial vulnerability can be a byproduct of networks' representational compression, rather than flaws in the learning process or non-robust inputs.