Enhancing Reliability of Distributed Learning over Edge Networks

Federated learning (FL) holds a captivating allure as it is capable of making the most of distributed data while safeguarding the privacy of users. However, to further propel the development and application of FL, numerous daunting challenges lie ahead that must be surmounted. These challenges encompass data heterogeneity, concerns regarding security, and the communication bottleneck that occur in edge Internet of Things (IoT) scenarios. In this paper, taking inspiration from the remarkable success of biological intelligence (BI) in the application of swarm organisms, we put forward a novel edge learning approach. This approach is achieved by integrating AI-enabled and BI-enabled optimization methods, i.e., stochastic gradient descent (SGD) and particle swarm optimization (PSO), which is named as link-reliable and Byzantine-robust distributed swarm learning (LB-DSL). By incorporating a very small quantity of globally shared data samples among local devices, the proposed LB-DSL can effectively alleviate the problems caused by nonindependent and identically distributed (non - i.i.d.) data as well as Byzantine attacks. Moreover, through a multi-user selection mechanism, our LB-DSL is able to counteract the instability of communication links. By comprehensively exploiting the swarm intelligence with the mechanism of exploration-exploitation, our LB-DSL is able to discover a more favorable global optimal solution. Consequently, it attains better learning performance when compared to the standard FL that merely utilizes AI-enabled SGD. The experimental results clearly demonstrate that our LB-DSL has a higher learning accuracy, shows greater reliability in terms of communication links, and exhibits stronger robustness against Byzantine attacks and non-i.i.d data.