Optimization of recurrent neural networks for high-performance intrusion detection in network traffic

Optimization of recurrent neural networks for high-performance intrusion detection in network traffic

연구 분야: Artificial Intelligence

학회: Cluster Computing

This paper presents a novel Optimized Recurrent Neural Network (O-RNN) framework for intrusion detection, designed to address the limitations of traditional neural network models in analyzing network traffic data. The O-RNN leverages advanced sequence reshaping techniques to convert network traffic into ordered temporal segments based on connection times, capturing complex temporal dependencies inherent in flow-based features, even when connections occur in parallel. This innovative approach allows for a more nuanced analysis of sequential data compared to conventional models. The O-RNN integrates adaptive Long Short-Term Memory (LSTM) layers that dynamically adjust to specific dataset characteristics, effectively capturing both short- and long-range dependencies while mitigating the vanishing gradient problem. This adaptability enables the O-RNN to model interdependencies between connections and temporal patterns, outperforming traditional RNNs, CNNs, and FNNs in complex and diverse intrusion detection tasks. Additionally, a hybrid optimization strategy extends beyond conventional Adam optimization by incorporating novel regularization and self-adjustment techniques, significantly improving training stability, reducing computational costs, and enhancing accuracy. Experimental results demonstrate that the O-RNN achieves superior computational efficiency across multiple datasets, with time costs of 334 on IoT-23, 432 on CICIDS2017, and 203 on KDD, outperforming CNN models, which exhibit significantly higher computational overheads of 4329 on IoT-23 and 4044 on KDD. LSTM variants achieve a moderate balance between time and accuracy, while FNNs, despite their lower time costs, fall short in capturing the intricate patterns required for effective intrusion detection. The O-RNN offers a scalable, efficient, and highly accurate solution for intrusion detection that addresses the evolving challenges of modern network environments.