SEAFL: Transforming Federated Learning for Enhanced Privacy in 6G-Enabled Vehicles

• Published in: 2023 Annual International Conference on Emerging Research Areas: International Conference on Intelligent Systems (AICERA/ICIS) pp. 1 - 8
• Main Authors: Vinita, L Jai, Vetriselvi, V
• Format: Conference Proceeding
• Language: English
• Published: IEEE 16.11.2023
• Subjects: 6G mobile communication; Data models; Differential privacy; Federated learning; Internet of Vehicles; Privacy; secured aggregation; Software-defined Vehicular fog computing; Training; Wearable devices
• DOI: 10.1109/AICERA/ICIS59538.2023.10420354

The utilization of artificial intelligence (AI) within the 6G-enabled Internet of Vehicles (IoV) has proven to be crucial in tackling complex issues, primarily through the adoption of deep learning methodologies. However, in certain sensitive data-driven sectors like healthcare and autopilot systems, traditional centralised training raises privacy concerns. To address this, federated learning (FL) has emerged as a popular approach, enabling collaborative model learning without exposing local data. However, recent investigations have highlighted that malevolent actors can exploit shared parameters to undermine critical applications such as autonomous driving systems, medical data from wearable devices, emergency message propagation scenarios, and the decision-making processes of industrial robots. In this article, we introduce a novel and robust aggregation mechanism named Secure and Efficient Aggregation for Federated Learning (SEAFL), tailored to the context of the 6G-enabled Internet of Vehicles (IoV). This approach aims to address the increasing privacy concerns and security threats arising in the collaborative learning landscape of the IoV, where sensitive data privacy is of paramount importance. Within the context of SEAFL, we amalgamated local differential privacy (LDP) into the realm of federated learning, employing a three-tier aggregation protocol within the software-defined vehicular fog (SDVF) architecture. Leveraging the incorporation of multiple aggregation levels and a privacy budget, SEAFL enhances privacy and accuracy by iteratively perturbing and aggregating gradients, resulting in heightened privacy protection, minimised gradient exposure, preserved data integrity, improved global model accuracy, and fortified resilience against adversarial threats. Through comprehensive experiments conducted on attack datasets, SEAFL outperforms alternative FL-based Local Differential Privacy (LDP) solutions, showcasing superior performance in terms of both accuracy and efficiency. This makes SEAFL a promising approach for secure and efficient federated learning in sensitive vehicular applications like emergency message dissemination scenarios.