2DSig-Detect: A semi-supervised framework for anomaly detection on image data using 2D-signatures

The rapid and widespread deployment of AI within critical systems raises serious questions about their security against adversarial attacks. Models performing image-related tasks, are vulnerable to integrity violations causing misclassifications that do not compromise normal system operation, but rather, produce undesirable outcomes for targeted inputs. This paper introduces a novel technique for anomaly detection in images called 2DSig-Detect , which is a 2D-signature-embedded semi-supervised framework rooted in rough path theory. We demonstrate that 2D-signatures can be applied to detect both adversarial examples and backdoored data, mitigating against test-time and training-time integrity attacks. Our results demonstrate both the efficacy and superior computational efficiency of our 2D-signature method for detecting adversarial manipulations to the training and test data.