acceptodds
Desk rejected at ICLR 2027

MoE-TS-IDS: A Mixture-of-Experts Time Series Model for Intrusion Detection in Medical IoT Networks

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Internet of Medical Things (IoMT) networks face severe intrusion threats where heterogeneous traffic exhibits multi-scale temporal patterns under class imbalance. Existing single-model IDS fail to simultaneously capture local, sequential, and global flow features. We propose MoE-TS-IDS, a Mixture-of-Experts Time Series IDS that serializes traffic features intotime series, encodes them via a BiLSTM encoder, and routes them through three parallel expert networks (CNN, LSTM, Transformer) with a learnable gating network and multi-head attention fusion. On WUSTL-EHMS-2020 (3-class),MoE-TS-IDS achieves 99.97% accuracy and AUC 1.0000.On CICIDS2017 (binary), it achieves 98.08% Macro F1 and AUC 0.9997, surpassing all deep learning baselines by +2.43% Macro F1. Ablation studies confirm contributions of the BiLSTM encoder (−1.93%), dynamic gating (−1.32%), and attention fusion (−0.95%). Gating weight analysis reveals polarized routing: normal traffic favors LSTM (g2=0.906) and attack traffic favors Transformer (g3=0.898), demonstrating interpretable autonomous expert specialization.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
IOT networks, MOE, Instrusion detection

Related papers

Discussion (0)

Sign in to comment.