Smart home networks consist of heterogeneous devices, fragmented protocols, and weakly consistent security policies, making household gateways vulnerable to device takeover, botnet expansion, credential abuse, and abnormal control behaviors. To address the joint constraints of detection accuracy, local deployment, and response timeliness, this paper proposes an embedded artificial intelligence-based cyber protection scheme for smart home networks, named EAI-SHGuard. The scheme unifies traffic statistics, device profiles, and state deviation signals into window-level security objects, and introduces a dual-branch lightweight detector that combines depthwise temporal convolution, linear attention, context-gated fusion, prototype regularization, and quantization-aware training. A risk grading module, furthermore, maps the results of detection into classified local processing measures, which include logging, speed restriction and isolation processing. Experiments were conducted on CICIoT2023, N-BaIoT, and ToN_IoT using standardized preprocessing and embedded deployment metrics. Under this protocol, EAI-SHGuard achieved Macro-F1 scores of 99.13%, 98.41%, and 97.84% on the three datasets, with an average of 98.46%. The model size is reduced to 3.6 MB, and the single-sample inference latency on a Raspberry Pi 4B is 6.1 ms. Case analysis further demonstrates that the proposed scheme can mitigate Mirai-like camera outbreaks and brute-force smart lock attacks through differentiated local responses. The results indicate that embedded AI can provide a practical security layer for smart home gateways when model compactness, contextual awareness, and policy linkage are considered together.
