Backdoor Attack Detection in Decision Tree Classifiers Using SHAP Explainability

Backdoor attacks pose a critical threat to the integrity and reliability of machine learning systems, including decision tree (DT) classifiers, which are widely employed in various domains. Detecting such attacks is crucial for maintaining trust in these systems. This paper presents a novel approach for enhancing the detection of backdoor attacks in DT classifiers leveraging Shapley Additive Explanations (SHAP). By examining the contribution of individual features to model predictions, SHAP provides valuable insights into model behavior, aiding in the identification of anomalous patterns indicative of backdoor attacks. For the first time in the literature, we propose a methodology that integrates SHAP explainability alongside decision path tracing into the detection process, enabling the identification of backdoor-induced biases in DT classifiers. Experimental results on public datasets for hardware Trojans and microscopy images of circuits demonstrate the effectiveness of the proposed approach in accurately detecting backdoor attacks.