Improving The OPENFL Mechanism to Avoid Federated Learning Poisoning Attacks in Blockchain Networks

Abstract

The integration of federated learning with blockchains represents a promising paradigm for collaborative artificial intelligence systems that preserve privacy. However, these decentralized environments are vulnerable to poisoning attacks, where malicious participants can sabotage the final model by sending malicious updates. This research paper proposes an enhanced mechanism to counter Random Noise Attack within blockchain-based federated learning environments. The enhanced mechanism is based on the OPENFL mechanism architecture, with an improvement to the method of making the voting decision inside it, by replacing the evaluation based on model accuracy and validation data with cosine similarity that measures the alignment of weight directions. We compared the proposed mechanism with the original OPENFL based on validation data, FEDCHALLENGER which uses cosine similarity, and FEDRDF which relies on Fast Fourier Transform, based on several metrics: efficiency, time cost, and GAS COST. Our experiments, conducted on a local Ethereum network using the Ganache simulator, the MNIST dataset, and a Convolutional Neural Network model, showed that the enhanced mechanism-maintained efficiency and improved time cost by 62.9% compared to the original OPENFL without causing an increase in cost (GAS) on the Ethereum network.