TSM-KFE: A Two-stage-shared Federated Learning Model Based on Key Feature Extraction

Yuanyuan Zhang, Jingwen Liu, Hanyang Zhang, Jingpeng Li, Liangyin Chen, Yanru Chen

Abstract: To address the issue of data heterogeneity and data privacy in blockchain and federated learning-based data-sharing models, this paper proposes a Two-stage-shared Federated Learning Model based on Key Feature Extraction (TSM-KFE). Specifically, in the TSM-KFE data-sharing model, data providers use Variational Autoencoders (VAE) to extract key features from their local data and share these key features before the federated learning task starts, thereby mitigating data heterogeneity among participating nodes. In addition, TSM-KFE employs blockchain and differential privacy technologies during the key feature-sharing stage to enhance data security. The experimental results show that our model improves by 20.35% compared to the baseline methods on the SVHN dataset with $a$ = 0.05. Furthermore, experimental results on different data sets and different heterogeneity conditions show that our model has higher compatibility and superior performance than the latest existing research results.