EASE-CloudNet: Adaptive Safety Alignment for Edge–Cloud SLMs via GNN-Informed Selective Reasoning and Multi-Objective Distillation

Abstract

EASE-CloudNet is a two-phase safety-alignment framework for generative small language models (SLMs) deployed on resource-constrained edge nodes. Its input is a natural-language user query and its output is a safe, helpful natural-language response or refusal; network-traffic classification and resource-scheduling actions are outside the task evaluated in this study. In Phase 1, a cloud teacher uses a security policy graph to generate structured safety rationales and response targets, which are distilled into Qwen2.5-1.5B/3B and Llama3.2-3B students. In Phase 2, an offline heterogeneous graph and a two-layer GraphSAGE model identify vulnerable semantic regions; these vulnerability targets supervise a lightweight edge-side router. We formulate deployment cost as a differentiable gate-conditioned expectation, so measured latency and energy constants affect the router through its reasoning probability. In the Qwen2.5-1.5B ablation experiments, the full model obtains 3.9% StrongREJECT ASR, 54.7% MMLU accuracy, and 70 average generated tokens; an A100 reference profile reports 18.8 ms/query and 2.37 J/query, or 3.3% latency and 2.6% measured GPU-energy overhead over the unaligned model. Physical edge runs measured a direct/reasoning end-to-end latency of 41.2/68.7 ms on Jetson Orin NX and 62.5/105.3 ms on Snapdragon 8 Gen 3, with a direct/reasoning energy of 0.48/0.79 and 0.71/1.18 J/query, respectively. Equal-seed Holm–Bonferroni-corrected tests confirm lower ASR than EASE on StrongREJECT and WildJailbreak for all three base models (p<0.01).