X-HDNN: Explainable Hybrid DNN for Industrial Internet of Things Backdoor Attack Detection
Love Allen Chijioke Ahakonye, Cosmas Ifeanyi Nwakanma, Jae Min Lee, Dong‐Seong Kim · 2023
This study proposes a hybrid deep neural network (HDNN) framework, called X-HDNN, for detecting backdoor attacks in Industrial Internet of Things (IIoT) data. The X-HDNN combines LeakyReLU and focal loss functions to reduce false positives and losses. A comparative analysis of the performance of the primary deep neural network and the proposed X-HDNN had remarkable improvement in f-score value from 57% to 78% and loss of 0.0044 to 0.0014. It also incorporates the SHAP explainability technique to provide interpretable and reliable detection. Evaluating the X-HDNN model using backdoor impact ratio, feature importance score, and decision confidence helps understand the model’s outcomes and the significance of each feature. The findings enhance trust in the model and facilitate better decision-making based on the provided explanations.