Unlocking potential of open source model training in decentralized federated learning environment
Ekaterina Pavlova, Grigorii Melnikov, Yury Yanovich, Alexey Frolov · Blockchain Research and Applications · 2025
The field of Artificial Intelligence (AI) is rapidly evolving, creating a demand for sophisticated models that rely on substantial data and computational resources for training. However, the high costs associated with training these models have limited accessibility, leading to concerns about transparency, biases, and hidden agendas within AI systems. As AI becomes more integrated into governmental services and the pursuit of Artificial General Intelligence (AGI) advances, the necessity for transparent and reliable AI models becomes increasingly critical. Decentralized Federated Learning (DFL) offers decentralized approaches to model training while safeguarding data privacy and ensuring resilience against adversarial participants. Nonetheless, the guarantees provided are not absolute, and even open-weight AI models do not qualify as truly open source. This paper suggests using blockchain technology, smart contracts, and publicly verifiable secret sharing in DFL environments to bolster trust, cooperation, and transparency in model training processes. Our numerical experiments illustrate that the overhead required to offer robust assurances to all peers regarding the correctness of the training process is relatively small. By incorporating these tools, participants can trust that trained models adhere to specified procedures, addressing accountability issues within AI systems and promoting the development of more ethical and dependable applications of AI.