Federated Learning for Object Detection in Autonomous Vehicles

Abstract

With the recent proliferation of Artificial Intelligence (AI), object detection is becoming increasingly ubiquitous. It is one of the key features of Autonomous Driving Systems. In current applications, object detection models are usually trained at a centralized location by collecting data from multiple sources. This raises concerns about data privacy among other issues. This paper addresses data privacy through a Federated Learning (FL) approach. FL architecture aims at preserving data privacy while maintaining performance by training the model in a decentralized manner. In this paper, we analyze how FL impacts the performance of object detection in a real-world traffic environment. We have constructed a prototype FL system and evaluated it on the KITTI Vision Benchmark 2D image dataset. In our prototype, object detection models are trained locally on a vehicle’s dataset, and the resultant weights are securely aggregated, using symmetric encryption techniques during data transfer, at the global server to yield an improved model. The FL model converged at 68% mean average precision. We compared the performance of object detection using FL to the traditional deep learning approach and noticed significant difference between the two models.