Adversarial Attack and Disturbance Detection by Hadamard-Coded Output Representations for Object Detection and Semantic Segmentation
Researchers have developed a new framework called HadamardNet to improve the robustness of object detection and semantic segmentation models against adversarial attacks. This framework utilizes Hadamard-coded output representations, which offer better calibration and allow for more effective detection of disturbances compared to traditional one-hot encodings. The novel approach includes an optimized decoding procedure and a method to exploit prediction inconsistencies for enhanced security. Evaluations show HadamardNet achieves state-of-the-art performance in detecting perturbations while maintaining competitive accuracy on clean data. AI
IMPACT Enhances AI model security by providing better detection of adversarial attacks and disturbances.