Intrusion Detection with Uncertainty based Loss Optimized Multi-Task Learning

Among all cybersecurity techniques, intrusion detection systems are the first and effective defense line for network and system. However, to develop an accurate, robust intrusion detection system with decent generalizability, challenges like limited labeled data, balancing between detection rates and false alarms need to be overcome. We propose an effective and efficient approach, in which the multi-class intrusion classification was performed by Multi-task Learning (MTL). A hard parameter sharing MTL model is utilized for the multi-class attack detection, in which uncertainty-based loss optimization is utilized to boost model performance and compute the best weights for each task. Comparisons of the proposed approach with single task learning (STL) models are conducted, and the results validate its balanced performance on different traffic. Moreover, when it comes to the capability to identify rare intrusions with limited samples, MTL model can out-perform other powerful STL models like DNN, CNN, RNN and LSTM.