Facial recognition technology integrated into IoT systems shows significant potential in device authentication, secure access control, and intelligent scenario interaction, but ensuring the confidentiality of biometric data such as facial features remains a key challenge. This paper proposes an IoT privacy protection framework that combines CNN-based facial recognition with homomorphic encryption and distributed edge computing. It uses MTCNN for facial alignment, FaceNet to extract 512-dimensional biometric embedding vectors, and CKKS homomorphic encryption to achieve end-to-end encrypted processing of facial features. The framework leverages edge computing to reduce latency in real-time IoT scenarios while meeting device privacy compliance requirements. Experiments on the LFW dataset show an identification accuracy rate of 98.06%, with encrypted computation performance comparable to that of plaintext, meeting the practical application needs of IoT devices and enhancing the robustness of facial recognition.