Personalized Heterogeneous Federated Learning with Gradient Similarity

In the conventional federated learning (FL), the local models of multiple clients are trained independently by their privacy data, and the center server generates the shared global model by aggregating local models. However, the global model often fails to adapt to each client due to statistical and systems heterogeneities, such as non-IID data and inconsistencies in clients' hardware and bandwidth. To address these problems, we propose the Subclass Personalized FL (SPFL) algorithm for non-IID data in synchronous FL and the Personalized Leap Gradient Approximation (PLGA) algorithm for the asynchronous FL. In SPFL, the server uses the Softmax Normalized Gradient Similarity (SNGS) to weight the relationship between clients, and sends the personalized global model to each client. In PLGA, the server also applies the SNGS to weight the relationship between client and itself, and uses the first-order Taylor expansion of gradient to approximate the model of the delayed clients. To the best of our knowledge, this is one of the few studies investigating explicitly on personalization in asynchronous FL. The stage strategy of ResNet is further applied to improve the performance of FL. The experimental results show that (1) in synchronous FL, the SPFL algorithm used on non-IID data outperforms the vanilla FedAvg, PerFedAvg, and FedUpdate algorithms, improving the accuracy by $1.81\!\sim\!18.46\%$ on four datasets (CIFAR10, CIFAR100, MNIST, EMNIST), while still maintaining the state of the art performance on IID data; (2) in asynchronous FL, compared with the vanilla FedAvg, PerFedAvg, and FedAsync algorithms, the PLGA algorithm improves the accuracy by $0.23\!\sim\!12.63\%$ on the same four datasets of non-IID data.