On December 30, Professor Kim Mi-ran at the Department of Mathematics at Hanyang University announced that she has developed a ‘homomorphic encryption-based security technology’ that allows Reinforcement Learning(RL) to be performed safely while preventing personal information leaks.

As AI technology rapidly expands into fields handling sensitive data—such as healthcare, finance, and autonomous systems—the importance of security technology that can guarantee data confidentiality throughout the entire deep learning training process is growing. However, existing security technologies faced limitations in being applied to actual deep learning training due to the computational burden and numerical instability in the learning process.

In particular, Reinforcement Learning is structured such that information regarding ‘state,’ ‘action,’ and ‘reward’ is repeatedly accumulated, and neural network parameters are continuously updated. The learning process itself has been pointed out as a potential risk factor for personal information leaks. Nevertheless, existing homomorphic encryption methods were difficult to be applied effectively to iterative learning environments like RL due to high computational costs.

To resolve the problem, Professor Kim’s research team presented a new framework capable of performing the entire training process of deep learning-based RL directly on encrypted data. The team designed the system to allow forward and backward propagation operations of neural networks within a homomorphic encryption environment by applying ciphertext packing operation structures and polynomial approximation techniques. Furthermore, to improve the optimization process previously difficult in homomorphic encryption environments, they reconfigured the Adam optimization algorithm, widely used in deep learning, to fit the encrypted environments.

In this way, they effectively reduced computational bottlenecks and training instabilities caused by high-order nonlinear operations. They successfully performed all training stages, including model parameter updates, in an encrypted state for the Soft Actor-Critic(SAC)-based RL algorithm. Using the framework, not only the state, action, and reward information but also the intermediate results generated during training are processed by an external computing server without decryption.

To verify the effectiveness of the proposed framework, the research team conducted RL experiments in an encrypted environment. The results showed that encrypted deep 토토사이트 하피-based RL maintained performance degradation within 10% compared to non-encrypted environments while securing stable 토토사이트 하피 convergence.

Professor Kim explained, "This research is an example showing that homomorphic encryption technology can go beyond simply protecting AI inference to safe actual learning process of deep learning models. It can be utilized as a reliable learning infrastructure across AI services handling sensitive data."

This research was conducted with support from the National Research Foundation of Korea(NRF) and the U.S. National Institutes of Health(NIH). The findings were published on December 1 in the international AI journal Nature Machine Intelligence (IF 23.9).

In the paper, Empowering artificial intelligence with homomorphic encryption for secure deep reinforcement 토토사이트 하피, Chi-Hieu Nguyen (Doctoral researcher at University of 토토사이트 하피 Sydney) served as the first author. Professor Thai Hoang Dinh, Professor Diep N. Nguyen, Dr. Kristin Lauter(Senior Director at FAIR Labs North America), and Professor Kim Mi-ran(Hanyang University) participated as corresponding authors.