Optimized compressed sensing for communication efficient federated learning

In recent years, data privacy preservation has received increased attention in artificial intelligence. Federated learning, as a paradigm for privacy-preserving machine learning, can considerably reduce the risk of privacy leakage by training models on local data. In federated learning, however, the clients and server must interact twice in each round of federated training, consuming abundant communication resources. To address the above issue, this work proposes an enhanced compressed sensing federated learning algorithm, which compresses and reconstructs the local network models trained on the clients using compressed sensing. To enhance the accuracy of the reconstructed models, we optimize the measurement matrix in compressed sensing using a genetic algorithm. In addition, we suggest an interleaving training and reconstruction method to improve the learning performance of the compressed models. Through a large number of experiments, we demonstrate that the proposed method is capable of maintaining high learning accuracy while accomplishing a large compression ratio for deep network models in federated learning. & COPY; 2023 Published by Elsevier B.V.