As various artificial intelligence (AI)-related services emerge, attempts to apply AI models to industrial and service fields are ongoing. In preparation for the increase in demand for such AI services, the importance of data required to create AI models is increasing. As data for applying services is a key element in training AI models and improving performance, the privacy and security of data have also recently emerged as significant issues. Although various AI learning platforms exist, users must share data that they have personally created and refined. Consequently, users who do not want to share personal data tend to be reluctant to access and use these platforms. Therefore, in this paper, we built a platform that can create and share specific AI models without sharing data. We designed and built a platform that can perform blockchain-based federated learning to ensure the authenticity and privacy of the user’s learning model and global model, allowing the learning of the final result model according to the sharing model of each dataset during each learning session. Each user learns based on their personal dataset, and the weights are integrated based on the learned models to create a single global model. The contribution of each user is measured based on the blockchain-based model creation, leading to the development of a high-performance AI model. We conducted experiments using the MNIST and COVID-19 datasets, focusing on data independence. The results showed that significant results could be achieved with just 10–20% data sharing. These experimental results confirmed that federated learning is possible in an environment where data independence is maintained and individual users do not share data.
Kwon et al. (Fri,) studied this question.