Februus: Input Purification Defense Against Trojan Attacks on Deep Neural Network Systems

We propose Februus; a novel idea to neutralize insidous and highly potent Trojan attacks on Deep Neural Network (DNN) systems at run-time. In Trojan attacks, an adversary activates a backdoor crafted in a deep neural network model using a secret trigger, a Trojan, applied to any input to alter the model's decision to a target prediction---a target determined by and only known to the attacker. Februus sanitizes the incoming input by devising an extraction method to surgically remove the potential trigger artifacts and use an inpainting method we propose for restoring the input for the classification task. Through extensive experiments, we demonstrate the efficacy of Februus against backdoor attacks, including advance variants and adaptive attacks, across vision applications. Notably, in contrast to existing approaches, our approach removes the need for ground-truth labelled data or anomaly detection methods for Trojan detection or retraining a model or prior knowledge of an attack. We achieve dramatic reductions in the attack success rates; from 100% to 0.25% (in the worst case) with no loss of performance for benign or trojaned inputs sanitized by Februus. To the best of our knowledge, this is the first backdoor defense method for operation in black-box setting capable of sanitizing trojaned inputs without requiring costly labelled data.