DLD: Defending Against Backdoors Using Differential Logits

The existing research has shown that deep learning models are vulnerable to backdoor attacks, in which malicious actors can manipulate the model’s predictions by embedding backdoors. As the stealthiness of these backdoor trigger mechanisms increases, traditional methods for detecting backdoors become increasingly inadequate. This paper presents a novel approach for detecting backdoors based on analyzing variations in model logits. Since the primary objective of a backdoor is to alter the model’s prediction results, detecting backdoors through the analysis of logits variations offers a robust method for backdoor determination. To improve the accuracy of identifying backdoor models, our proposed method not only examines the logits variations of the model under investigation but also introduces an auxiliary normal model with a simpler structure. The presence of a backdoor in the model under investigation is determined by contrasting the differences in logits variations between the model under investigation and the normal model. Furthermore, to enhance detection speed, we introduce a dataset reduction mechanism to streamline the testing dataset. Experimental results demonstrate that our method successfully defends against high-stealth backdoor attacks such as WaNet and LIRA. Moreover, in terms of detection speed, our method outperforms mainstream methods like Neural Cleanse, which takes several hours for a single detection, by requiring only several minutes.