UTS researchers debut privacy-preserving AI framework for encrypted data

Overview

On 18 March 2026, researchers at the University of Technology Sydney (UTS) announced a world-first privacy-preserving artificial intelligence framework that resolves one of the most persistent bottlenecks in modern technology adoption. Published in the prestigious journal Nature Machine Intelligence (2026), the breakthrough introduces a deep reinforcement learning (DRL) system that utilises Fully Homomorphic Encryption (FHE) to process and analyse information. This development allows machine learning models to train on and generate decisions from sensitive datasets without ever decrypting or exposing the underlying raw data. For professional services, corporate legal counsel, municipal planners, and infrastructure developers, this technology offers a secure pathway to harness advanced predictive algorithms without compromising proprietary intellectual property or violating strict privacy mandates.

The integration of artificial intelligence in high-stakes professional sectors has historically been restricted by data security liabilities. Traditional machine learning models require data to be decrypted into plaintext before processing, which exposes sensitive information to potential cyber threats, third-party hosting vulnerabilities, and regulatory breaches. This limitation has forced environmental consultants, planning authorities, and corporate entities to keep highly valuable datasets siloed. Consequently, the industry has struggled to utilise the full predictive power of artificial intelligence for complex tasks like risk forecasting, geological modelling, and long-term project planning.

By enabling artificial intelligence to operate securely on encrypted information, the UTS framework dismantles the traditional trade-off between computational capability and data sovereignty. It ensures that data remains protected throughout its entire lifecycle: during transmission, during model training, and during decision generation. This is particularly relevant for Australian industries navigating complex regulatory landscapes, where the secure management of proprietary, environmental, and financial data is critical to maintaining market integrity and compliance.

Key details

The technical core of the UTS research addresses the inherent mathematical challenges of performing complex neural network calculations on encrypted data. Fully Homomorphic Encryption (FHE) is a cryptographic method that allows mathematical operations to be performed on ciphertexts, producing an encrypted result that, when decrypted, perfectly matches the outcome of operations performed on the original plaintext. While FHE has existed in various forms, applying it to deep reinforcement learning (DRL) has historically been considered computationally prohibitive. DRL models rely on continuous feedback loops, dynamic parameter adjustments, and iterative optimisation, which create an immense mathematical overhead when executed in an encrypted state.

The breakthrough achieved by the UTS research team solves this computational bottleneck through the development of a homomorphic encryption-compatible Adam optimiser. The Adam (Adaptive Moment Estimation) optimiser is a widely utilised algorithm for gradient-based optimisation of stochastic objective functions in neural networks. By successfully re-engineering this optimiser to function entirely within the encrypted domain, the researchers have eliminated the need to decrypt data during the model training phase. This ensures that the training operations, gradient calculations, and weight updates occur without exposing any plaintext values, preserving absolute data integrity.

Under this new framework, the operational workflow is fully decentralised to protect the user’s data privacy. The data owner encrypts their sensitive information locally using their private cryptographic keys before transmitting it to the cloud-hosted AI system. The AI model processes this encrypted data, executes the deep reinforcement learning algorithms, and generates decisions or predictive outputs that also remain fully encrypted. These encrypted results are then sent back to the local user, who is the sole holder of the decryption key. Consequently, the external AI service provider or cloud host never gains access to the raw data or the final decisions in plaintext, effectively eliminating the risk of data leaks or insider exploitation.

The research demonstrates that this framework can be scaled across various complex decision-making environments. The UTS team noted applications ranging from autonomous driving systems, where real-time navigation decisions must remain secure from external interception, to the back-end processing of Generative AI platforms. By proving that deep reinforcement learning can be executed with high accuracy and efficiency on homomorphically encrypted datasets, the study establishes a new benchmark for secure, collaborative machine learning across highly regulated industries.

Australian context

The development of a privacy-preserving AI framework carries profound implications for the Australian business and professional services landscape, particularly regarding data governance and regulatory compliance. In Australia, the handling of sensitive corporate, environmental, and personal information is strictly regulated under the Privacy Act 1988, the Australian Privacy Principles (APPs), and various state-level data protection acts. As organisations increasingly look to modernise their operations with machine learning, they must navigate the legal risks associated with sending proprietary or client-confidential data to third-party cloud servers. The UTS framework provides a technical solution that aligns with these strict statutory requirements, allowing firms to adopt advanced AI tools while maintaining total data sovereignty.

This technology is highly relevant to collaborative projects across the environmental, planning, and infrastructure sectors, where multiple stakeholders must share modelling inputs, site assessments, and forecasting data without exposing commercially sensitive or confidential records. By allowing encrypted datasets to be processed by shared AI systems, the framework opens the door for joint risk assessments, cross-jurisdictional planning analyses, and secure participation in industry data pools, supporting more informed decision-making while keeping proprietary information firmly under the control of its owner.

References and related sources

• Primary source: english.news.cn
• NEPM Assessment of Site Contamination

How iEnvi can help

iEnvi provides specialist consulting services relevant to this topic. Our team includes CEnvP Site Contamination Specialists with experience across contaminated land, groundwater, remediation, ecology, and regulatory compliance.

• iEnvi contaminated land investigation services
• iEnvi remediation and validation services
• iEnvi expert services and independent review services

---

This is an iEnvi Machete news summary. Prepared by iEnvi to summarise the source article for contaminated land, groundwater, remediation, approvals and site risk professionals.

Published: 17 Jun 2026

Need advice on this topic? Speak to an iEnvi expert at info@ienvi.com.au or 1300 043 684, or contact us online.