Desilo, a private AI company specialising in fully homomorphic encryption (FHE), said on Oct. 16 it will take part in a project titled “Development of physical AI-based multi-layer security technology for collaborative intelligence systems” under the “Collaborative intelligence physical AI-based software platform R&D ecosystem” programme supported by South Korea's Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA).

In the project, Desilo will be responsible for developing end-to-end encryption based on fully homomorphic encryption (FHE) and privacy-preserving federated learning technology. The task is led by Raonsecure. Senstone, Tilon, Desilo, IT Station, Sogang University Industry-Academic Cooperation Foundation and Jeonbuk National University Industry-Academic Cooperation Foundation will jointly develop security technologies to protect overall physical AI operations.

The consortium will link AI model safety verification, identity and access management for robots and AI agents, protection of data confidentiality and computational integrity, a secure AI development environment and operational technology (OT) network security. It plans to build a multi-layer security infrastructure to develop and operate physical AI safely.

A key element is applying secure aggregation to federated learning. Federated learning is a method in which each organisation trains a model in its own environment using its data and pools the resulting model updates to reflect them in a shared model. Desilo plans to develop technology to encrypt each organisation’s model updates for transmission and aggregate them without decrypting individual updates.

This will lay the groundwork for participating organisations to jointly advance AI without directly sharing raw data. The goal is to protect companies’ core intellectual property, such as process data and model weights, by reducing the risk that manufacturing know-how and training information for individual models could be exposed during training.

Lee Seung-myeong (이승명), CEO of Desilo, said, “For AI collaboration on manufacturing sites to expand, companies must be able to jointly train models while protecting accumulated process data and know-how.” He added, “Through this task, we will develop fully homomorphic encryption-based data protection and secure federated learning technology to build a foundation that allows companies to use AI while safeguarding the confidentiality of core data.”