A two-stage federated optimization algorithm for privacy computing in Internet of Things

With the advent of the Internet of things (IoT) era, federated learning plays an important role in breaking through traditional data barriers and effectively realizing data privacy and security in the process of sharing. However, the demand of the practical problems makes the algorithm still have great challenges in effectively balancing various factors, such as privacy security, accuracy, computing efficiency and so on. To challenge this problem, a two-stage federated optimization algorithm based on robust and multitasking learning is designed. In optimization client local model stage, an adaptive weight assignment mechanism is adopted to guide robust learning based on multiple untrusted sources data, which aims to ensure the credibility and robustness of the client model and obtain a reliable client model. To address the information leakage problem during the server-client global model optimization stage, a privacy patch layer is added to the client local model in multitask learning and maintain its privacy parameters stored on the client model during the global model parameter aggregation process, which aims to meet the personalized requirements and performance requirements of protecting the model privacy. To effectively measure the performance of our algorithm, two extensive experiments are carried out to verify the robustness and accuracy of model under different datasets, respectively. In addition, simulation results show that our algorithm successfully suppresses the impact of corrupted or irrelevant sources on performance, and its performance is better than the other two robust distributed learning baseline methods in the client local model optimization stage. At the same time, our algorithm achieves better accuracy performance than other advanced personalized optimization algorithms in the server-client global model optimization stage. Finally, it achieves a good balance between robustness, computational efficiency and model privacy protection.(c) 2023 Elsevier B.V. All rights reserved.