Privacy-Preserving Edge Video Analytics

Abstract

Edge video analytics (EVA) emerges as a promising paradigm to decentralize video analytics tasks to the edge of the network and thus improve user experience by enabling faster task execution for users. However, it is challenging to enable edge-based model training while still satisfying the requirement of user privacy. A faster model training process is commonly at the cost of more disclosure of user privacy. We propose a federated learning driven privacy-preserving model training framework called FedEVA for edge video analytics, which can protect user privacy and ensure a fast convergence rate. Instead of directly sending gradients to the parameter server, users conduct a local perturbation operation on users’ update information, and then send the perturbed gradients to the parameter server. The parameter server at the edge can update the neural network model directly over the perturbed gradients. We carefully design the perturbation function to conceal partial information about data while efficiently performing computation over the gradients from multiple users. Different from crypto-based methods, our perturbation process is a light-weight operation. We conduct extensive evaluations using large-scale real-world datasets to verify the effectiveness of our FedEVA framework and compare with other baseline algorithms. The results show that our FedEVA framework can improve the degree of privacy preservation, and still maintain the same level of convergence rate.