Trojan cleansing with neural collapse

Abstract Trojan attacks are sophisticated training-time attacks on neural networks that embed backdoor triggers that force the network to produce a specific output on any input that includes the trigger. With the increasing relevance of deep networks that are too large to train with personal resources and that are trained on data too large to thoroughly audit, these training-time attacks pose a significant risk. In this work, we connect trojan attacks to Neural Collapse, a phenomenon wherein the final feature representations of over-parameterized neural networks converge to a simple geometric structure. We provide experimental evidence that trojan attacks disrupt this convergence for a variety of datasets and architectures. We then use this disruption to design a lightweight, broadly generalizable mechanism for cleansing trojan attacks from a wide variety of different network architectures and experimentally demonstrate its efficacy.