Transformer-Encoder Detector Module: Using Context to Improve Robustness to Adversarial Attacks on Object Detection

Deep neural network approaches have demonstrated high performance in object\nrecognition (CNN) and detection (Faster-RCNN) tasks, but experiments have shown\nthat such architectures are vulnerable to adversarial attacks (FFF, UAP): low\namplitude perturbations, barely perceptible by the human eye, can lead to a\ndrastic reduction in labeling performance. This article proposes a new context\nmodule, called \\textit{Transformer-Encoder Detector Module}, that can be\napplied to an object detector to (i) improve the labeling of object instances;\nand (ii) improve the detector's robustness to adversarial attacks. The proposed\nmodel achieves higher mAP, F1 scores and AUC average score of up to 13\\%\ncompared to the baseline Faster-RCNN detector, and an mAP score 8 points higher\non images subjected to FFF or UAP attacks due to the inclusion of both\ncontextual and visual features extracted from scene and encoded into the model.\nThe result demonstrates that a simple ad-hoc context module can improve the\nreliability of object detectors significantly.\n