A Multi-Objective AutoML-based Efficient Intrusion Detection System for EV Charging Networks

This paper addresses the challenge of deploying Intrusion Detection Systems (IDS) in resource-constrained Electric Vehicle Charging Systems (EVCS). Traditional machine learning-based IDSs often prioritize detection accuracy at the expense of computational efficiency, making them unsuitable for edge devices with limited memory and processing power. To solve this, the author introduces a Multi-Objective Automated Machine Learning (MOO-AutoML) framework that treats detection performance, inference latency, and model size as competing objectives to be optimized simultaneously.

The proposed framework utilizes a two-stage optimization process. First, it employs a LightGBM-based automated feature selection (AutoFS) method that ranks features by importance and selects a compact subset based on an accumulated importance threshold. Second, it uses the Non-dominated Sorting Genetic Algorithm III (NSGA-III) to search for the optimal combination of feature selection thresholds and LightGBM hyperparameters. By using a small 2% subset of the training data for this search, the framework significantly reduces the computational cost of the AutoML process while maintaining high-quality results on the full test set.

Experimental results on the CICEVSE2024 and CICIDS2017 datasets demonstrate that the MOO-AutoML approach successfully generates a Pareto front of models. These models provide flexible trade-offs between accuracy and deployment constraints, consistently achieving lower 99th percentile (P99) inference latency and smaller model sizes than state-of-the-art baseline models. The use of P99 latency ensures that the system remains responsive even under worst-case traffic conditions, which is critical for real-time security in cyber-physical systems.

As EV charging infrastructure becomes increasingly integrated with IoT and smart grid technologies, the attack surface for cyber threats grows. This research provides a practical, automated path for security administrators to deploy robust intrusion detection models that are tailored to the specific hardware limitations of charging stations, ensuring that security does not come at the cost of system performance or availability.