Abstract
The rapid proliferation of Internet of Things (IoT) devices has introduced substantial challenges for network security analysis, particularly under encrypted communication environments where payload-based inspection becomes ineffective. Traffic fingerprinting based on lightweight behavioral metadata provides a promising alternative; however, existing approaches still suffer from two critical limitations: insufficient modeling of heterogeneous traffic sequences and excessive computational requirements that hinder edge deployment. To address these challenges, this paper proposes a lightweight multi-task traffic fingerprinting and autonomous anomaly isolation framework for IoT-oriented edge scenarios. The proposed method employs a payload-agnostic spatiotemporal representation pipeline that extracts directional packet-level metadata, including packet lengths and temporal intervals, from the initial handshake sequence (first N = 15 packets), without accessing application-layer payloads. A shared lightweight backbone combining one-dimensional convolution, bidirectional recurrent modeling, and self-attention is designed with an Asymmetric Dual-Head Perceptron (ADHP) to jointly perform device identification and intrusion detection while mitigating task interference. Furthermore, a hierarchical dual-temperature knowledge distillation strategy is introduced to transfer discriminative knowledge from a cloud-scale teacher model to an edge-oriented student model. Experimental evaluations on a CPU-only edge-emulation platform demonstrate that, compared with the uncompressed cloud-scale teacher model (1,569,764 parameters, 9.45 MFLOPs), the proposed student model reduces the parameter size by 91.20% (down to 138,116 parameters, 1.57 MFLOPs, and 0.53 MB storage) while maintaining an attack detection F1-score of 99.95% and a Top-1 device identification accuracy of 64.65% across 31 device classes, with a single-core inference latency of only 0.484 ms. Notably, the accuracy gap relative to session-complete feature engineering baselines is attributable to the strict payload-agnostic, short-sequence operating constraint rather than model capacity—a deliberate design trade-off for privacy compliance and sub-millisecond inference. The proposed framework also enables a lightweight detection–scoring–isolation mechanism with 12.4 ms end-to-end response capability, indicating its potential for practical edge-oriented IoT security applications.