Adaptive Model Transfers and Aggregations for Efficient Federated Learning in IoT Edge Systems with Non-IID Data
Jenn-Wei Lin, Taewoon Kim, Hung-Jen Tu, Po-Hsien Kuo · 2023
Federated learning (FL) can associate with edge computing system to perform distributed and privacy-preserving learning for intelligent IoT applications. With individually collecting the data in each FL node, the training data may exhibit non-independent and identically distributed (non-IID) property, which may seriously impair the FL performance. To mitigate the impact of non-IID training data, we propose a virtual non-IID transformation approach by combing local datasets among FL nodes to form an IID dataset group for tuning each local training model. First, each local dataset is examined and the data label distribution is uploaded. Then, a perfect data label distribution is made as the best IID benchmark. For each communication round of FL, the local training model is verified to find lacking data label distributions. In next several rounds, the locally trained model will be sent to assistant FL nodes for compensating its uneven data label distribution based on the perfect IID benchmark. Thereafter all FL participating clients join the parameter aggregation at the end of each round same as the traditional FL approaches. The selection of assistant nodes will be transformed into the minimum-cost maximum-matches (MCMM) problem, and then we use integer linear programming (ILP) to obtain the optimal solution. Finally, we perform extensive simulation experiments to evaluate the effectiveness of the proposed approach.