ALERT: A lightweight defense mechanism for enhancing DNN robustness against T-BFA

DNNs have become pervasive in many security–critical scenarios such as autonomous vehicles and medical diagnoses. Recent studies reveal the susceptibility of DNNs to various adversarial attacks, among which weight Bit-Flip Attacks (BFA) is emerging as a significant security concern. Moreover, Targeted Bit-Flip Attacks (T-BFA), as a novel variant of BFA, can stealthily alter specific source–target classifications while preserving accurate classifications of non-target classes, posing a more severe threat. However, due to the inadequate consideration for T-BFA’s “targeted” characteristic, existing defense mechanisms tend to perform over-protection/-modification to the network, leading to significant defense overheads or non-negligible DNN accuracy reduction.

In this work, we propose ALERT, A Lightweight defense mechanism for Enhancing DNN Robustness against T-BFA while maintaining network accuracy. Firstly, fully understanding the key factors that dominate the misclassification among source–target class pairs, we propose a Source-Target-Aware Searching (STAS) method to accurately identify the vulnerable weights under T-BFA. Secondly, leveraging the intrinsic redundancy characteristic of DNNs, we propose a weight random switch mechanism to reduce the exposure of vulnerable weights, thereby weakening the expected impact of T-BFA. Striking a delicate balance between enhancing robustness and preserving network accuracy, we develop a metric to meticulously select candidate weights. Finally, to further enhance the DNN robustness, we present a lightweight runtime monitoring mechanism for detecting T-BFA through weight signature verification, and dynamically optimize the weight random switch strategy accordingly. Evaluation results demonstrate that our proposed method effectively enhances the robustness of DNNs against T-BFA while maintaining network accuracy. Compared with the baseline, our method can tolerate $6.7\times$ more flipped bits with negligible accuracy loss ($<0.1\text{\%}$ in ResNet-50).