Adversarial Patch Detection: Leveraging Depth Contrast for Enhanced Threat Visibility

Neural networks have proven to be extremely effective at tasks such as image classification and object detection. However, their security and robustness are controversial. Even state-of-the-art object detectors can be fooled by localized patch attacks, which might lead to safety-critical incidents. In these attacks, adversaries place a subtle adversarial patch in an image, causing detectors to either miss real objects or detect phantom objects. These adversarial patches often force state-of-the-art detectors to make highly confident but incorrect predictions. The practical implications of these attacks in real-world settings further increase the concern. This paper presents a unique method for detecting real-world adversarial patches using entropy-sensitive depth estimation. Therefore, we take advantage of the fact that adversarial patches typically introduce high local entropy and are located in front of an object. We have fine-tuned a monocular depth estimation neural network to exploit these two features to extract adversarial patches from an image. Using this approach, we are able to achieve a true positive detection rate of 77.5% on the APRICOT test set.