Century Communities Explores AI Privacy Advances with DESILO's Revolutionary THOR Framework
- DESILO's THOR framework enhances AI privacy by enabling large language model inference under homomorphic encryption.
- The framework achieves significant performance improvements, allowing secure AI use without compromising sensitive data.
- DESILO aims to lead in privacy-focused AI solutions, collaborating on initiatives like the Harvest™ platform for secure data analysis.
Groundbreaking Advances in AI Privacy with DESILO's THOR Framework
In a notable development within the artificial intelligence sector, DESILO Inc. introduces its innovative THOR framework, which enables large language model (LLM) inference to operate under homomorphic encryption. This technology significantly enhances privacy in AI applications, addressing a critical concern in the industry. The unveiling of this framework marks a pivotal moment in privacy-preserving research, as the company collaborates with Professor Miran Kim's team from Hanyang University, leading to a joint research paper accepted for presentation at ACM CCS 2025. This prestigious conference is recognized for its focus on computer and communication security, underscoring the importance of DESILO's contributions to the field.
The THOR framework achieves two remarkable milestones: the ability to operate a widely-used open-source model without any retraining while maintaining the integrity of homomorphic encryption. Additionally, it boasts near practical runtime performance for processing inputs, achieving deployment-relevant speeds on a single GPU. The framework demonstrates significant enhancements in performance, achieving a 5.3 times improvement for plaintext to ciphertext operations and a staggering 9.7 times for ciphertext to ciphertext operations. These advancements set new benchmarks in the realm of secure AI, allowing organizations to leverage LLMs without compromising sensitive data.
Seungmyung Lee, CEO of DESILO, emphasizes the importance of this achievement in advancing privacy-preserving AI. He highlights the company's commitment to developing solutions that prioritize data security, such as the forthcoming Harvest™ platform, which aims to facilitate secure and privacy-safe data analysis across various institutions. By prioritizing trusted and privacy-centric applications, DESILO positions itself as a leader at the intersection of AI technology and data security, paving the way for future innovations.
In addition to the THOR framework's capabilities, this development is poised to influence industries that require stringent data protection measures, such as healthcare and finance. As organizations increasingly seek to integrate AI while safeguarding user privacy, DESILO's advancements could serve as a model for responsible AI deployment. The collaboration with Cornami further underscores the importance of partnerships in enhancing the efficacy and reliability of homomorphic encryption technologies.
The implications of DESILO’s work extend beyond its immediate applications, potentially reshaping how data privacy is approached in AI. As the demand for secure AI solutions grows, the THOR framework could lead to widespread adoption of homomorphic encryption, establishing new standards in the industry and fostering trust among users and institutions alike.