A New Approach for Low-Latency, High-Accuracy Anomaly Detection at the Edge: Benchmarking Quantized Autoencoders, LSTMs, and Lightweight Transformers on RT-IoT2022 Time-Series Traffic

Abstract

This study benchmarks edge-optimized deep learning models for real-time anomaly detection in resource-constrained IoT environments using the RT-IoT2022 dataset, which includes four benign protocols and nine cyberattack types. Three architectures a quantized autoencoder (QAE), compact LSTM, and lightweight Transformer were deployed on a Raspberry Pi 4 and evaluated on F1-score, latency, model size, and energy per inference. The QAE achieved optimal performance with 98.7% F1- core, 142 KB memory footprint, 1.8 ms latency, and 4.2 mJ energy consumption, outperforming alternatives under strict edge constraints. While the LSTM showed better recall on rare attacks and the Transformer captured long-range dependencies at higher computational cost, the QAE delivered the best overall trade-off for deployable security. The work reframes model selection around hardware- ware co-design rather than architectural complexity, demonstrating that intelligently compressed, reconstruction-based approaches surpass heavier models in efficiency and effectiveness. Findings provide a reproducible framework for low-latency, privacy-preserving intrusion detection in smart healthcare and industrial IoT, advocating a paradigm shift toward minimal sufficiency over maximal capacity in edge AI design.