Hierarchical Local Sensitivity Analysis Method for Optimization of Deep Learning-based Anomaly Detection in Communication Network

• Published in: 2022 IEEE/CIC International Conference on Communications in China (ICCC) pp. 451 - 456
• Main Authors: Yue, Gang, Sun, Zhuo, Fan, Jinpo
• Format: Conference Proceeding
• Language: English
• Published: IEEE 11.08.2022
• Subjects: Analytical models; anomaly detection; Communication network; Computational modeling; Deep learning; Integrated circuits; local sensitivity analysis; Neurons; Robustness; robustness improvement; Sensitivity analysis; structure compression
• DOI: 10.1109/ICCC55456.2022.9880845

Deep learning-based anomaly detection techniques in communication network have been widely researched and generally viewed as a classification problem based on the benign data baseline. However, the model deployment on edge devices is urgent to be optimized as compact structure with robust detection capability due to the constraints of computational and storage capacity. This paper proposed a hierarchical local sensitivity analysis (H-LSA) method to quantify the model's sensitive correlations by deriving the "neuron-layer-model" sensitivity matrices. The method provides a perspective on the sensitivity propagation from the basic unit of the deep learning model, i.e., neuron to the overall model. Furthermore, we demonstrate the applications of the hierarchical sensitivity matrices on CNN-based anomaly detection in the industrial control network communication, which shows the remarkable optimization capabilities of structure compression and robustness improvement.