Reproduce “CAG: A Real-time Low-cost Enhanced-robustness High-transferability Content-aware Adversarial Attack Generator” on OpenResearch

About this paper

Huy Phan, Yi Xie, Siyu Liao, Jie Chen, Bo Yuan

A Content-aware Adversarial Attack Generator (CAG) is introduced, enabling real-time generation of adversarial examples with significantly reduced memory costs for targeted attacks. The method achieves over a 500-fold speedup compared to iterative attacks and reduces memory requirements from gigabytes to megabytes, while enhancing robustness against defenses and demonstrating high transferability across diverse DNN classifiers.

Key insights

• Encoding class information into a compact, generalizable embedding within a single generative model drastically reduces memory overhead for targeted attacks across numerous classes.
• By leveraging Class Activation Maps, adversarial perturbations can be strategically placed on content-aware regions, making the attacks more resilient to common defense mechanisms.
• Introducing stochasticity through dropout in the perturbation generation process encourages the model to learn more generalized attack patterns, leading to higher transferability in black-box settings.

Summary by alphaXiv.