Edge Federated Learning with Blockchain for Private Model Updates

The patent segments the centralized training process into distributed local training at edge nodes. Each node trains models locally on its own data, keeping data private while contributing only model updates to the global training process. This segmentation resolves the contradiction by enabling accurate model training without centralizing sensitive data.

The patent introduces blockchain as an intermediary layer that mediates between edge nodes and the global model aggregation. The blockchain securely transmits and validates model updates without exposing raw data, acting as a trusted intermediary that preserves data privacy while enabling collaborative training for improved model accuracy.

The patent merges multiple federated learning approaches by combining blockchain-based secure transmission, Kalman filter-based weight adjustment, and containerized deployment. This integration resolves the accuracy limitation of traditional federated learning by adding sophisticated weight adjustment mechanisms and ensuring reliable model aggregation across heterogeneous edge nodes.

The patent dynamically adjusts model training parameters using Kalman filters, which optimize weight updates based on observed performance and uncertainty. This parameter adaptation enables the system to achieve higher model accuracy while maintaining the data privacy benefits of federated learning, resolving the trade-off between privacy preservation and model performance.

The patent implements periodic model update transmission instead of continuous updates. Edge nodes transmit model updates at optimized intervals determined by the blockchain network state and performance requirements. This periodic action maintains model adaptability while significantly reducing communication overhead and energy consumption compared to frequent continuous updates.

The patent allows each edge node to maintain local model quality and adaptability independently while contributing to the global model. Nodes can perform local adaptations without immediate central coordination, reducing the frequency and volume of communications required while preserving model adaptability to local conditions.