A Secure Object Detection Technique for Intelligent Transportation Systems

Abstract

Federated Learning is a decentralized machine learning technique that creates a global model by aggregating local models from multiple edge devices without a need to access the local data. However, due to the distributed nature of federated learning, there is a larger attack surface, making cyber-attack detection and defense challenging. Although prior works developed various defense strategies to address security issues in federated learning settings, most approaches fail to mitigate cyber-attacks due to the diverse characteristics of the attack, edge devices, and data distribution. To address this issue, this paper develops a hybrid privacy-preserving algorithm to safeguard federated learning methods against malicious attacks in Intelligent Transportation Systems, considering object detection as a downstream machine learning task. This algorithm involves the edge devices (e.g., autonomous vehicles) and road side units to collaboratively train their model while maintaining the privacy of their respective data. Furthermore, this hybrid algorithm provides robust security against data poisoning-based model replacement and inference attacks throughout the training phase. We evaluated our model using the CIFAR10 and LISA traffic light dataset, demonstrating its ability to mitigate malicious attacks with minimal impact on the performance of main tasks.