Optimization of User Resources in Federated Learning for Urban Sensing Applications

Sparsh Gupta, Ayshika Kapoor, Dheeraj Kumar

Keywords: Federated learning, Urban sensing, Noise map, Secure multiparty computation, Turbo-Aggregate, Resource-adaptive

TL;DR: A resource adaptive, efficient, and privacy preserving secure multiparty computation scheme for FL-based urban sensing systems where different participants have varying computation and communication resources.

Abstract:

Participants involved in federated learning based urban sensing tasks are prone to unknowingly leak sensitive information to the central server or an adversary. Secure multiparty computation is a promising solution for protection against inference attacks without compromising on application model accuracy. However, existing secure multiparty computation protocols are resource intensive as they require multiple communication and computation rounds between participants and the central server. For urban sensing applications, where application model is usually a spatiotemporal map over a large area, this is an even more challenging problem to deal with. To achieve real time sensing using frequent model updates, existing state-of-the-art secure multiparty computation protocols such as Turbo-Aggregate shall not be feasible.
This paper presents an optimised Turbo-Aggregate protocol, we call Resource Adaptive (ReAd) Turbo-Aggregate, which is a secure multiparty computation scheme designed specifically for urban sensing applications where different participants have varying computation and communication resources. It is an adaptive scheme which grants the flexibility of modifying the space and time granularity of the application model in each round of aggregation to suit the participants’ network, processing, and battery resources while retaining the features and security of the parent Turbo-Aggregate protocol. The proposed approach is verified with simulation experiments for the noise mapping task of urban sensing applications. The results demonstrate that the proposed solution provides a useful application model along with the benefits of user privacy using limited computation and communication resources of participating users.