BFL-SE: Blockchain federated learning with secure and effective weight-assignment model aggregation

With the proliferation of interactive displays in edge–cloud environments, including emerging AR/VR visualization platforms, ensuring secure visual data processing while maintaining real-time responsiveness has become a critical requirement. Federated learning emerges as a new way of edge–cloud computing framework to handle privacy and security issues in display-centric edge computing. However, federated learning still suffers from problems such as over-reliance on central servers and the tampering of models. Blockchain technology provides decentralized and tamper-proof ability in the field of finance and has its potential in edge–cloud computing. In this paper, we propose a framework BFL-SE that combines blockchain and federated learning. We design several modules to further improve the framework’s security and efficiency. For security, we integrate FLTrust into our proposed framework and compute the trust scores of the clients to filter out malicious clients. For efficiency, we predict the client’s future model loss reductions using the client’s historical loss values to identify clients with better performance. The trust scores and the loss predictions constitute the aggregation weights. The final obtained global model is uploaded to the blockchain. The experiment results show that our framework balances the security and efficiency of the FL process. The model accuracies under label parameter attacks are all greater than 85%, and the convergence is faster than the baselines.