Adversarial Risk and the Dangers of Evaluating Against Weak Attacks
This work addresses a critical problem for AI security researchers by exposing vulnerabilities in existing adversarial defense evaluations, highlighting that incremental improvements may be misleading without rigorous testing against stronger attacks.
The paper investigates the evaluation of adversarial robustness in machine learning models, revealing that current defenses may only appear robust because they optimize against weak attacks rather than true worst-case adversarial risk. By developing new gradient-free attacks, the authors show that several recent defenses can be broken, reducing their accuracy to near zero.
This paper investigates recently proposed approaches for defending against adversarial examples and evaluating adversarial robustness. We motivate 'adversarial risk' as an objective for achieving models robust to worst-case inputs. We then frame commonly used attacks and evaluation metrics as defining a tractable surrogate objective to the true adversarial risk. This suggests that models may optimize this surrogate rather than the true adversarial risk. We formalize this notion as 'obscurity to an adversary,' and develop tools and heuristics for identifying obscured models and designing transparent models. We demonstrate that this is a significant problem in practice by repurposing gradient-free optimization techniques into adversarial attacks, which we use to decrease the accuracy of several recently proposed defenses to near zero. Our hope is that our formulations and results will help researchers to develop more powerful defenses.