UNIST Prevents Personal Data Breaches Amid Rapid AI Development with Federated Learning

Privacy-Preserving AI Federated Learning Technology... High Performance Across Diverse Data Distributions

FedGF, the IT Industry's Focus: Enhancing User Data Protection and AI Performance Simultaneously

As AI continues to advance, concerns over user privacy violations are becoming more serious, and a core federated learning technology has emerged to address these issues.

This is expected to benefit on-device AI learning, which is drawing attention from IT companies.

A research team led by Professor Yoon Sunghwan at the Graduate School of Artificial Intelligence, UNIST (President Park Jongrae), has developed a technology called FedGF (Federated Learning for Global Flatness) that can enhance AI performance while protecting user privacy. It is expected to resolve the issue of user data leakage.

From the top, Professor Yoon Sunghwan, Researcher Lee Taehwan.

The research team developed a method that consistently delivers high performance even in various user data distribution scenarios. Previous technologies only showed excellent performance in environments similar to the user data distribution, but performed poorly in different environments.

Federated learning protects personal information by training deep learning models on user devices, but data differences have limited its performance. FedGF achieves high accuracy by creating optimized models through locally trained models on each device, without sending data to a central server.

FedGF is also highly efficient. Complete training is possible with less communication resources compared to existing methods. This is especially advantageous for mobile devices that use wireless communication such as Wi-Fi.

Comparison between the existing method (left) and the proposed algorithm (right).

Professor Yoon Sunghwan said, "Federated learning technology will be a key stepping stone to solving AI privacy issues," and added, "It will greatly help IT big tech companies overcome privacy problems and heterogeneity in distributed data."

First author Lee Taehwan, a researcher, added, "With FedGF technology, companies can obtain high-performance AI models without privacy violations, and it will play a major role in various fields such as IT, healthcare, and autonomous driving."

The research results were published online on July 20 at the world-renowned international conference ICML (International Conference on Machine Learning). The research was supported by the Ministry of Science and ICT's International Collaborative Research on Information Security and the ICT Broadcasting Innovation Talent Development Project.