#South Korea #Seoul #Desilo #THOR Framework #Privacy AI

DESILO's Groundbreaking THOR Framework

In a significant leap for privacy-safe artificial intelligence (AI), DESILO Inc. has unveiled its innovative THOR framework, which enables large language models (LLMs) to function entirely under homomorphic encryption. This pioneering approach is not just a theoretical advancement; it is recognized through a joint research paper that has been accepted for presentation at the highly esteemed ACM CCS 2025 security conference.

Revolutionizing Privacy in AI

Homomorphic encryption allows computations to be carried out on encrypted data, thus protecting sensitive information throughout the entire inference process. This level of privacy is essential in today’s data-driven world, where concerns over unauthorized access and data breaches are escalating. With THOR, DESILO has managed to eliminate the need for retraining existing models, making it the first framework to operate a complete LLM under such encryption, achieving near practical performance speeds.

Key Achievements

The groundbreaking research conducted in collaboration with Professor Miran Kim's team at Hanyang University has demonstrated two critical achievements:

1. Execution of an Open-Source Model: The researchers successfully operated a widely-used open-source LLM under homomorphic encryption without requiring any retraining. This was achieved by replacing conventional computations with encryption methods, ensuring that both inputs and outputs remained encrypted.

2. Enhanced Runtime Performance: The results were striking, with the model processing around 128 tokens (approximately two sentences) on a single GPU at speeds relevant for deployment. Notably, core matrix multiplication performance improved remarkably: a 5.3x increase for operations on plaintext to ciphertext and a staggering 9.7x improvement for ciphertext to ciphertext operations.

These outcomes set a new standard in the realm of privacy-preserving AI solutions, showcasing the capability of DESILO's THOR to deliver reliable and rapid inference processes while safeguarding data integrity.

Future Implications

Seungmyung Lee, CEO of DESILO, remarked on this achievement's significance, noting, "This CCS acceptance signals a groundbreaking advance in executing large language models under homomorphic encryption, which is a pivotal academic milestone in the realm of privacy-preserving AI research." Lee also emphasized the dual initiative at DESILO to advance homomorphic encryption research in collaboration with partners like Cornami while also fast-tracking product development aimed at integrating trusted Privacy AI into everyday application scenarios.

The THOR framework lays the foundation for more secure future solutions from DESILO, like Harvest™, which aims to facilitate secure and privacy-centric analysis shared across various institutions. This innovation not only strengthens data security but also opens new avenues for cooperation in fields that heavily rely on data sharing, such as healthcare and finance.

Conclusion

As AI continues to evolve, ensuring data privacy becomes paramount. DESILO’s THOR framework is at the forefront of this evolution, providing a robust solution that meets the growing demands for safe data practices without sacrificing performance. This dual commitment to research and real-world application demonstrates DESILO's leadership in the future of AI technology.

In this fast-paced landscape of technological advancements, the implications of DESILO's breakthroughs will be closely watched as they pave the way towards genuinely private and secure AI solutions.