Edge-Prioritize IDS: Zero-Retraining Class Prioritization for Real-Time Edge Intrusion Detection

Deploying deep neural networks-based intrusion detection systems on resource-constrained edge devices demands inference strategies that balance latency, energy, and accuracy under shifting threat landscapes. This paper presents Edge-Prioritize IDS, a class-prioritized early-exit framework that accelerates inference for high-risk attack classes without post-deployment retraining. A lightweight K-dimensional control vector encodes per-class runtime priorities and steers samples toward earlier exits via adaptive normalization and cost-sensitive training. Evaluation across five benchmarks NSL-KDD, CIC-IDS2017, UNSW-NB15, WISDM, and CIFAR-10 on an NVIDIA Jetson TX2 shows that Edge-Prioritize IDS preserves baseline accuracy (up to 99.6%) while reducing latency by up to 55% and energy by up to 50% for prioritized classes. Ablation studies isolate each component’s contribution, and a controlled distribution-shift experiment demonstrates the sliding-window heuristic’s ability to recover near-baseline latency within 500 samples under synthetic class-frequency drift. Once trained under the proposed framework, the model requires no additional retraining, firmware updates, or additional memory beyond the priority vector itself when runtime priorities change.