Morphological real-time defense against physical attacks on convolutional neural networks

This article presents an effective algorithm for defending object detection systems against physical visual attacks, based on analyzing morphological differences between video stream frames. The proposed method, independent of the neural network architecture used, combines mosaic approximation of image shape, binarization of morphological changes, local filtering, and re-verification of suspicious areas using a neural network model. The system successfully detects the presence of adversarial patches disguised as clothing or scene elements. Experiments show that implementing the proposed algorithm significantly improves the 𝐹1-metric for human detection, reduces the number of false positives, and ensures stable real-time performance. The algorithm was tested on various attack types (AdvTshirt, AdvTexture, NatPatch, AdvCaT) and detector models (YOLO, Faster R-CNN, RetinaNet), demonstrating generalization ability and practical applicability.