Developing a blockchain-based framework for decentralised federated edge learning

Abstract

The introduction of the federated learning technique has brought about many benefits from disrupting the traditional centralised process of machine learning. Privacy can be preserved by avoiding the concentration of full data sets on a single central server, which is instead now distributed amongst a group of model training worker nodes. However, there still exists potential security risks when the central coordinating server is compromised, misleading worker nodes when data corruption occurs or leading to loss of privacy through speculative data reconstruction attacks. To overcome these issues, a blockchain-based framework is developed to serve as the data coordination backbone layer between the worker nodes and central model aggregator. This is achieved by using a blockchain to store tamper-proof records of model update data, both for individual workers as well as for the aggregated global model. These records point to the corresponding location of model update data uploaded onto the InterPlanetary File System (IPFS), overcoming the potential security and corruption risks of storing all data on a single central server. This framework is extensible on both ends of the federated learning process pipeline. On the input side, any arbitrary type of model data can be supported since there are no specific data structure requirements imposed. As for the output, the model record metadata stored on the blockchain can be adapted for use in higher-level applications such as cross-chain data trading marketplaces.