HIPFL: A Hierarchical Incentive-Based Personalized Federated Learning

Federated learning is a distributed collaborative training approach that aims to balance model performance and privacy. In this distributed environment, clients with high-quality data can be incentivized to participate more in the training, thus improving the performance of federated learning. But if this process performs in a Non-IID environment, the value of the model is an important metric to determine the level of client participation. Therefore, an effective evaluation mechanism is needed. However, when each client evaluates values, there will be a risk of privacy leakage which will prevent the client from actively participating in the session. To solve these challenges, in this paper, we propose hierarchical incentive-based personalized federated learning (HIPFL) based on privacy model value evaluation and reverse weight updating. This method aims to achieve secure, fair and personalized federated incentivization. Through extensive experiments on three Non-IID datasets, the HIPFL improves the average accuracy by 8.54% compared to existing four federated learning incentivization methods and effectively resists attacks from malicious nodes. Compared with existing models that use encryption for value evaluation, the HIPFL has an average accuracy improvement of 13.53%.