UTS Researchers Build AI Privacy System That Learns From Fully Encrypted Data

Researchers at the University of Technology Sydney have developed what they describe as the first deep reinforcement learning system capable of learning and making decisions while the data it processes remains fully encrypted.

The framework, published in Nature Machine Intelligence, uses fully homomorphic encryption (FHE) to allow AI models to operate on sensitive data without ever decrypting it. The work was carried out in collaboration with Meta AI Research and South Korea’s Hanyang University.

The Problem With AI And Private Data

Deep reinforcement learning (DRL) underpins a range of technologies, from autonomous vehicles to the decision-making processes behind generative AI. However, DRL systems typically require access to raw data in order to learn, which can expose personal, financial or commercially sensitive information to external platforms.

This tension between AI capability and data privacy has been a persistent challenge for the field.

Associate Professor Hoang Dinh, who led the project, outlined the significance of the work.

“For years, the field has struggled to reconcile AI with strict privacy requirements,” Dinh stated. “Our work shows AI can now learn and make decisions while keeping sensitive data completely protected.”

How The System Works

In the proposed system, a user encrypts all of their information before sharing it with an external AI agent. The AI then learns how to make decisions directly from the encrypted data, without access to the original information.

The decisions the AI produces also remain encrypted. They are only decrypted and applied by the user locally. This cycle repeats until the system reaches the desired performance level, with privacy maintained throughout the entire process.

The central technical challenge the team had to overcome involved the mathematical operations used in modern AI training. Functions such as inverse square roots, which are critical to optimisation algorithms, are difficult to compute on encrypted data and typically fail under homomorphic encryption.

The UTS team addressed this by developing an HE-compatible Adam optimiser that bypasses the need for high-degree polynomial approximations of inverse square roots on encrypted data.

Performance Close To Unencrypted Models

Early results indicate the encrypted DRL model performs within 10 per cent accuracy of standard unencrypted techniques, while maintaining full data confidentiality.

Associate Professor Diep Nguyen, a co-researcher on the project, pointed to the broader implications.

“This could be a cornerstone for the next generation of AI, including generative AI,” Nguyen commented.

A Cross-Border Research Effort

The project was carried out through the Australia Vietnam Strategic Technologies Centre, bringing together UTS, Meta and Hanyang University.

The UTS team was led by Associate Professor Hoang Dinh and Associate Professor Diep N. Nguyen, alongside PhD Student Hieu Nguyen. On the international side, Dr Kristin Lauter, Director of Research Science at Meta AI Research’s North American Labs, and Associate Professor Miran Kim from Hanyang University contributed to the work.

Dinh described the collaboration as central to the outcome.

“By collaborating with partners like Meta and Hanyang University, we were able to combine theoretical innovation with real-world insights,” he explained. “It’s this kind of international partnership that positions UTS at the forefront of privacy-preserving AI research.”

An Ethical Dimension

Dinh framed the research as part of a broader responsibility facing the AI community as the technology becomes more embedded in daily life.

“As AI becomes more embedded in our everyday lives, ensuring it can learn without compromising personal or organisational privacy isn’t just a technical challenge – it’s an ethical responsibility,” he noted.

“That’s why researchers are focusing not just on what AI can do, but how it can do it responsibly.”

Last Updated on March 17, 2026 by Nick Ross