Stealthy Adversarial Patch for Evading Object Detectors Based on Sensitivity Maps

Previous studies on physical adversarial attacks primarily concentrate on evaluating the effectiveness of the attacks, by enhancing the strength of perturbations to achieve their goals. However, this approach usually results in conspicuous and striking patterns that can be easily recognized by human observers. Moreover, the size of the adversarial patch is a critical factor for its stealthiness. Large scale patches are not only noticeable but also impractical for real-world applications, as they are easily detected and removed. In order to tackle this issue, we propose a novel approach to generate naturalistic adversarial patches for effective and robust attacks on object detection systems by leveraging sensitivity maps to identify and drop less significant pixel points of the patches, thus reducing their size and enhancing their invisibility. Extensive experiments demonstrate that our method generates patches that are less conspicuous compared to state-of-the-art techniques, while maintaining strong attack performance.