<p>The rapid growth of smart home ecosystems has introduced critical challenges in user and device authentication due to personalized behavioral patterns, heterogeneous data distributions, and exposure to adversarial threats. Conventional centralized authentication methods and standard federated learning frameworks face significant limitations in preserving privacy, managing nonindependent data, and sustaining real-time edge performance. This study presents PFL-FogAuth, a personalized and privacy-preserving federated learning framework for fog-assisted smart home authentication. The framework integrates behavior-aware modeling, asynchronous federated optimization, differential privacy, robust aggregation, and latency-aware scheduling within a unified architecture. Each fog node independently trains lightweight local models on private data and contributes differentially private updates to a global model, enabling secure, accurate, and adaptive authentication across distributed environments. Experimental evaluations on three multimodal and nonindependent datasets, including human activity recognition, WiFi Channel State Information, and DeepFake voice biometrics, demonstrate that PFL-FogAuth achieves up to 6.9% higher accuracy and 38% lower false acceptance rates compared with state-of-the-art methods, while meeting real-time inference constraints. Large-scale simulations of heterogeneous fog networks and distributed optimization under adversarial conditions were executed using high-performance computing resources to ensure scalability and reproducibility. This work enhances secure federated learning and establishes a practical foundation for integrating intelligent IoT authentication systems with supercomputing infrastructures.</p>