- Abstract: The non-robustness of neural network policies to adversarial examples poses a challenge for deep reinforcement learning. One natural approach to mitigate the impact of adversarial examples is to develop methods to detect when a given input is adversarial. In this work we introduce a novel approach for detecting adversarial examples that is computationally efficient, is agnostic to the method used to generate adversarial examples, and theoretically well-motivated. Our method is based on a measure of the local curvature of the neural network policy, which we show differs between adversarial and clean examples. We empirically demonstrate the effectiveness of our method in the Atari environment against a large set of state-of-the-art algorithms for generating adversarial examples. Furthermore, we exhibit the effectiveness of our detection algorithm with the presence of multiple strong detection-aware adversaries.
- Supplementary Material: zip