Federated Learning in Mobile Edge Computing: Optimization, Privacy, and Applications for Cybersecurity

Abstract

In the big data era, compared with traditional centralized machine learning, federated learning can greatly reduce data collection time, relieve computation burdens of local clients, and preserve data privacy. Recently, communication overhead in federated learning begins to attract increasing attentions. At first, federated learning needs several communication rounds between local devices and public servers for model aggregation, which are mostly through wireless networks. Besides, since uploaded model parameters in federated learning are still vulnerable to inference attacks, extra privacy masks are needed for them, such as Secret Sharing (SS) and secure Multi-Party Computation (MPC), which can further aggravate communication. In addition, the number of communication rounds can soar if adversaries try to slow convergence of federated learning by poisoning local data or uploaded model parameters. In this dissertation, we provide comprehensive analysis about optimizing communication overhead in federated learning from the above three aspects. In detail, we discuss the following four research projects: 1) We consider communication overhead reduction through convergence performance optimization in federated learning via introducing centralized machine learning-based adaptive learning strategies to the model parameter update rule. Convergence upper bounds under our optimization scheme are derived after each communication round with a certain number of local iterations, and after a given number of communication rounds. Through comparison with the bounds of original federated learning, we theoretically analyze how those strategies should be tuned to help federated learning effectively optimize convergence performance and reduce overall communication overhead; 2) We propose a privacy-preserving task scheduling strategy based on (2,2) SS and mobile edge computing to reduce data processing latency, in which locally-learned model parameters are separated into two portions before uploaded to public edge servers for parameter aggregation based on MPC. We show that the related privacy constraint can be enforced through constructing a pairwise Markov chain. We further formulate the whole task scheduling problem as a stochastic latency minimization problem and solve it by converting it into a linear programming problem; 3) We further extend the (2,2) case to (R,L) case, and propose a communication-aware secret share placement strategy to optimize communication overhead by minimizing weighted transmission hop counts in a hierarchical edge computing architecture. We show that the constructed optimization problem is NP-hard, and efficient heuristic algorithms can be applied to find sub-optimal solutions. We respectively evaluate two traditional heuristics, i.e. Genetic Algorithm (GA) and Particle Swarm Optimization (PSO), introduce two basic heuristics, i.e. top-down and bottom-up heuristic, and further propose an advanced algorithm, called Bottom-Up Top-Down (BUTD) heuristic. Based on comparison, we find that our proposed BUTD heuristic can outperform all the other four heuristics when communication among different shares of the same secret is comparable to that among different secrets; 4) Finally, we talk about combating data poisoning attacks through developing a federated self-learning intrusion detection system, which is based on a practical application of federated learning in cybersecurity. This application targets Controller Area Network (CAN bus) and is based on Graph Neural Network (GNN). We show that different driving scenarios and vehicle states will impact sequence patterns and data contents of CAN messages. In this case, we develop a federated learning architecture to accelerate the learning process while preserving data privacy.