Resolving Data Heterogeneity Through Personalized Federated Learning With Global Calibration

Federated learning (FL) is gaining momentum as a significant solution for the considerable data generated in modern society. FL enables participants to train a shared model collaboratively so that data distributed across many clients can be properly utilized. In addition, FL does not require the clients to send their raw data to the server, which can prevent personal data leakage. However, data statistics across clients vary greatly in practice, which is the problem of data heterogeneity. Data heterogeneity or non-I.I.D. data is a huge challenge in FL. The high degree of data heterogeneity will cause degradation of model accuracy. To address these issues, we propose Personalized Federated Learning with Global Calibration (PerFedGC). This approach is based on personalization and further uses our proposed calibration mechanism to improve performance. Personalized FL is a technique that allows each client to train its own model, which is more suitable for its unique task than a global model. Furthermore, we propose a calibration mechanism that assists participants in leveraging shared knowledge while retaining their own information. We present the experimental results on various datasets and different scenarios. The results show that the proposed PerFedGC outperforms other FL algorithms, demonstrating our method’s performance enhancement and robustness.