Robust Training on the Edge: Federated vs. Transfer Learning for Computer Vision in Intelligent Transportation Systems

Sergei Chuprov, Raman Zatsarenko, Dmitrii Korobeinikov, Leon Reznik

Abstract: The synergy between the Internet of Things and Edge AI is revolutionizing industries by enabling real-time data processing on devices at the network’s edge, like sensors in Intelligent Transportation Systems (ITSs). However, a key challenge arises when deploying foundation Machine Learning (ML) models, trained on high quality data, on these Edge AI systems often subjected to Data Quality (DQ) variations. In this paper, we address this challenge by leveraging Transfer Learning (TL) and Federated Learning (FL) as strategies to mitigate the impact of DQ variations on ML application performance. While these strategies were not originally designed for this purpose, our findings demonstrate that both TL and FL can effectively enhance the robustness of ML applications in ITS scenarios that involve running ML processes on edge devices. We showcase this through a real-world traffic sign detection application, analyzing how TL and FL can be employed to improve model robustness against variations in DQ typically encountered by edge devices in ITS. We found that when high-quality data only is available for re-training, FL with Geometric Median aggregation allowed to train models performing on average by 20% better than in the TL scenario. Our results demonstrated that employing Geometric Median aggregation in FL allowed to increase accuracy by 6.7% on average across the all the considered cases in comparison to Federated Average aggregation. Additionally, employing varying DQ for re-training helps to further enhance ML performance if the application’s operation scenario involves high dynamics in the quality of input data.