A Trustless Federated Framework for Decentralized and Confidential Deep Learning

Nowadays, deep learning models can be trained on large amounts of web data on power hungry servers and be deployment-ready for specific real-world applications. With a state-of-the-art model architecture and a large publicly available dataset for pre-training, convolutional neural network models can be further fine-tuned via transfer learning for a related task. Nonetheless, the training process required of privacy-sensitive applications needs to protect data confidentiality and simultaneously boost performance using limited training data. For such data-deprived and privacy-centric learning, we introduce a trustless federated learning framework that seamlessly integrates deep learning models from different edge nodes using a blockchain-based architecture. Our framework performs federated learning without the need of a central server by leveraging a smart contract blockchain platform with a distributed file system for model storage. Users can locally train on their data while routinely benefiting from an enhanced model obtained by merging the models from all users in a decentralized fashion. Most importantly, this framework is free of the potential single point of failure in centralized federated learning. Besides, our framework has in-built incentive mechanisms to prevent model corruption and temper bad actors. We tested our framework on various computer vision datasets. The experimental results show that the merged model accuracy is on-par compared to a centralized federated training setup. To the best of our knowledge, this work represents the first systematic attempt at building a blockchain-based federated deep learning framework for computer vision. The code is publicly made available at: https://github.com/s-elo/DNN-Blockchain.