Topological Anomalies for Trojaned Neural Networks detection

Deep neural networks (DNNs) are increasingly vulnerable to adversarial attacks, particularly Trojan attacks, which embed malicious behaviors triggered by specific inputs. Detecting these attacks is challenging due to subtle model manipulations and limited data access. Topological data analysis (TDA) has shown promise in identifying structural deviations within neural networks through Persistent Homology (PH). Prior work has leveraged PH to distinguish Trojaned models by summarizing topological features. We propose an enhanced detection framework that use persistence images as a features for classification. Our approach improves robustness and accuracy. Additionally, we introduce a randomized variant of the method that significantly reduces execution time while maintaining classification accuracy.