An Early Exit Deep Neural Network for Fast Inference Intrusion Detection

An Early Exit Deep Neural Network for Fast Inference Intrusion Detection

연구 분야: Safety

학회: SAC '25: Proceedings of the 40th ACM/SIGAPP Symposium on Applied Computing

Deep Neural Networks (DNN) are currently state-of-the-art in intrusion detection literature, where authors typically escalate the network parameters to pave the way for accuracy improvements. However, in addition to increasing the inference computational costs, this can also render them unsuitable for resource-constrained devices, given their limited memory and processing capabilities. This paper introduces a new early exit neural network for fast inference and reliable intrusion detection. Our proposal partitions the DNN utilized for intrusion detection into branches, with the objective of classifying the majority of samples on the earlier branches, thereby reducing inference costs. Challenging samples that reach the final DNN branch are subsequently classified using a reject option, improving system reliability. In addition, the branches and rejection thresholds are selected as a multi-objective optimization task. Experiments on a new dataset with over 8TB of data and a year-long real network traffic showed the proposal's feasibility. Our scheme reduces the average inference computational costs by up to 82% while decreasing the average error rates by up to 3.3.