LEQ: Energy-Efficient Early-Exit Inference for On-Device Security in Consumer Electronics — Artifacts

Real measurement data and reproduction scripts for the LEQ paper (IEEE Consumer Electronics Magazine, CEMAG-OA-0026.R1, round-2 major revision). A trained early-exit T5 encoder (6 layers) performs 8-class on-device network-attack classification; an entropy threshold drives both per-layer early exit and edge-cloud routing, with a 5-condition security gate against confidence-inflation. Includes: GB10 GPU full-test-set sweep; Raspberry Pi 5 PMIC-metered per-input runs (tau sweep, per-sample tail latencies, 5-seed reruns with mean/std/paired-Wilcoxon statistics); a CALM-inspired security-gate-disabled ablation baseline (Figure 4 accuracy-energy Pareto frontier); the fixed-depth ablation; and a new adversarial validation suite added in this revision: a black-box query-based confidence-inflation attack (default vs. CALM-inspired gate, at 60- and 150-query budgets, and at a stricter entropy threshold), a random-perturbation per-layer vulnerability diagnostic, a forced-exit-layer attack isolating layer susceptibility from exit routing, and a hardened gate configuration (attention-stability 0.98, patience 3) that cuts attack success rate from 25.0% to 4.0% at a quantified early-exit-rate cost. All harness and analysis scripts included so every number in the paper regenerates end to end.