Blockchain-Based Auditing and Traceability Framework for Federated Learning Models

Abstract

Federated Learning (FL) is one of the potentially useful paradigms of collaborative machine learning, where various clients compute their models independently on their private data, while preserving privacy.However, FL in particular suffers from transparency, auditing and traceability issues, one of the key requirements to ensure reliability and trust in the global model.This paper proposes a federated learning system auditing and traceability framework based on blockchain.Each client trains a CNN on their data and updates the model, those updates are securely hashed, written to a blockchain and attached to metadata including the identity of the client, local accuracy and time.It is a way of assurance and responsibility, in which all contribution of clients is monitored accurately while training process.The global model is trained using an iterative model convergence process called federated averaging (FedAvg) that merges the input of the clients without sacrificing privacy.Empirical experiments on MNIST data set demonstrate competitive classification accuracy and transparent and tamper-proof auditing of model updates.The findings suggest that using blockchain technology in federated learning not just boosts security and accountability, but also increases traceability, and is therefore applicable to real-world applications of decentralized AI where trust and reliability are paramount.