Causality-Guided Counterfactual Debiasing for Anomaly Detection of Cyber-Physical Systems

Machine learning has become a promising technology for anomaly detection of cyber-physical systems (CPSs). However, the trained anomaly detection models always suffer from bias due to the scarcity of anomaly data in CPSs and the biased data collection process, which may poison the models' generalization ability. Recent debiasing methods are proposed to deal with the bias via resampling the training dataset, reweighting during the training phase, or adjusting the classification threshold. However, they may lose valuable information, need extra knowledge of the models, or lead to overfitting. Especially, they lack a causal understanding of the debiasing process, so they cannot point out the source and propagation of the bias and, thus, cannot deal with it in an explainable way. In this article, we propose a counterfactual debiasing framework to mitigate the bias in a well-trained model. First, we formalize the model's training and inference processes using causal graphs. Thus, we can understand the source and propagation of the model's bias through causal inference. Then, we use counterfactual inference to estimate the bias's detrimental causal effect on the prediction and remove it from the total causal effect. Therefore, we can conduct unbiased inferences with a biased model. The proposed method can remove the bias in an explainable way by incorporating causal graphs. Comprehensive experiments are conducted on seven real-world CPS datasets, i.e., IDA, MFP, ACS, SPF, UNS, NSL, and ICS. The results demonstrate the effectiveness, compatibility, and unbiasedness of the proposed approach.