Risk-aware learning-enabled control for safety-critical federated systems: A privacy-preserving approach for CIoT malware defense

Learning-enabled systems are increasingly deployed in distributed, safety-critical environments where reliability must be maintained under uncertainty, heterogeneity, and adversarial behavior. This challenge is particularly evident in the Consumer Internet of Things (CIoT), where Android-powered consumer electronics (CE) collaboratively participate in malware detection while operating under strict resource and privacy constraints. Conventional machine learning and deep learning approaches for malware detection, validated primarily on static datasets, provide limited assurance when such systems operate in dynamic, real-world conditions with unreliable participants. To address these challenges, this paper proposes a risk-aware, security-aware fuzzy privacy-preserving federated learning (FPPFL) framework for Android malware classification on resource-limited CE devices. The federated learning process is treated as a dynamical learning system whose stability and safety must be actively regulated during operation. A Paillier homomorphic encryption scheme provides a privacy-preserving instrumentation layer for secure gradient aggregation across distributed clients, while a fuzzy risk assessment (FRA) mechanism functions as a real-time safety controller that evaluates the reliability of local updates before global aggregation. The proposed framework is evaluated on multiple benchmark Android malware datasets, including Malgenome, Drebin, Tunadromd , and Kronodroid . Experimental results show that the FPPFL approach not only preserves privacy but also improves robustness against unreliable or adversarial updates while maintaining superior detection accuracy compared with existing federated and centralized models. Furthermore, it demonstrates favorable trade-offs in latency and energy consumption, validating its suitability for edge deployment. By integrating lightweight cryptography with adaptive risk regulation, this work illustrates how learning-enabled control principles can be applied to cybersecurity systems, enabling safe, robust, and privacy-preserving collaborative learning in CIoT environments.