Work

Packet-Window Anomaly Detection at the Edge: Latency and Energy Trade-Offs of a Distilled Transformer

Year: 2026
Type: article
Abstract: Deploying Transformer-based anomaly detection at the network edge requires jointly configuring input representation and model execution under monitoring and resource constraints. We characterize the encoder–regressor component of a previously defined distilled MiniLM packet-window detector, linking token budget, packet-window stride, and device operating mode to detection quality, scoring capacity, and energy cost. Using 5GAD-2022 and NVIDIA Jetson Nano and Xavier NX, we characterize four token budgets across five device–power-mode configurations through detection metrics, mean inference latency from 14.2 to 466.6 ms, latency distributions, and internal input-rail energy from 0.265 to 2.840 J per evaluated window. The analysis reveals diminishing detection gains as token budget increases while execution cost continues to grow, making representation budget a practical deployment control. Power-mode scaling introduces a platform-dependent trade-off: faster execution increases energy per decision on Nano, whereas Xavier NX provides substantial latency reductions with comparatively small energy changes. We translate these measurements into operating envelopes and configuration-selection criteria that relate model scoring capacity to packet-window stride and monitoring demand. Detection quality supplies a task-specific reference; the separate supervised classification head is outside the execution boundary. The resulting analysis provides a quantitative framework for configuring a Transformer-based anomaly-sensing component within an edge monitoring architecture, identifying how representation and execution choices can be coordinated to satisfy application requirements.
Source: Sensors
Language: English
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7236800460
FWCI:
Cites: 29
Cited by:
Related to: 10