Toward Efficient Neural Networks Through Predictor-Assisted NSGA-III for Anomaly Traffic Detection of IoT

Abstract

Edge computing enhances intrusion detection by extending its reach to smaller Internet of Things (IoT) devices and edge nodes, improving real-time detection and data privacy. However, due to the limited processing power and storage of edge nodes, lightweight models with high detection accuracy are urgently needed. The Neural Architecture Search (NAS) technique based on multi-objective genetic algorithms can simultaneously balance model complexity and performance, thus automatically designing models for fast and accurate detection. However, NAS requires training each model from scratch during the optimization process to obtain its detection accuracy as one of the fitness evaluation criteria. To address this limitation, we propose an efficient predictor-assisted NSGA-III algorithm. It uses proxy models to swiftly predict architecture accuracy, eliminating the need for complete training and greatly improving optimization efficiency. Furthermore, we have designed an innovative search space that allows for the reduction of internal channels and output feature map dimensions within the model, resulting in the creation of lightweight models with minimal impact on classification performance. The proposed method is validated by searching a narrower Pareto-optimal model of the competitive F1 score of 95.17% with 38.10 MB FLOPs on the UNSW-NB15 dataset. By adding a predictor, the number of optimizations per iteration increased, leading to faster convergence. Additionally, when comparing the search spaces before and after our design in the most complex structure (with 7 cells), the model’s classification error rate increased from 4.31% to 4.36%, while the FLOPs decreased from 271.13MB to 186.9MB.