AI and Cryptography
Can AI strengthen existing cryptanalysis techniques? Can we leverage cryptography to secure AI systems?
Overview
Although connections between the fields of AI (machine learning) and cryptography have long been considered, only recently have the two disciplines come into close contact. Today’s powerful AI models have proven helpful in cryptanalysis, while cryptographic methods have been leveraged to make AI models secure and private. Yet, the intersection between these two fields remains relatively understudied.
The Argus Lab’s research on AI and cryptography focuses on two key directions: leveraging AI tools to improve cryptanalysis, focusing particularly on problems in post-quantum cryptography; and assessing the security of cryptographic methods proposed for use in AI. Already, Argus Lab researchers have created novel AI-powered attacks against Learning with Errors, a hard math problem used in the newly standardized CRYSTALS-KYBER system. Ongoing work from the lab includes the development of new AI techniques for cryptanalysis, deeper scrutiny of proposed cryptographic implementations, and the discovery of new ways to secure AI models via cryptographic methods.
Related Projects
- Novel AI paradigms for cryptanalysis - ongoing work.
- Security analysis of cryptographic methods for image authentication (e.g. C2PA) - ongoing work.
- Cryptanalysis of Learning with Errors with transformer models - see here, here, here, here, and here.
- Fundamental learning techniques for AI-powered cryptanalysis - see here.
- Other AI-powered attacks on Learning with Errors - see here.
- Benchmarking AI and classical attacks on Learning with Errors - see here.