Google DeepMind Mathematically Verifies Key Elements of Signal

Google DeepMind Mathematically Verifies Key Elements of Signal

Researchers at Google DeepMind have used an agentic AI to mathematically prove several key cryptographic guarantees of the Signal protocol, according to Wired, which reported on the work. This demonstration, conducted as a proof-of-concept project, represents a notable advance for the widely used cryptography as well as for artificial intelligence itself.

The project lasted four weeks. It first provides new independent validation of the cryptography that protects a vast amount of communications on platforms like the Signal app, Google Messages, and WhatsApp. It also showcases Google DeepMind’s formal verification capabilities, thanks to a new AI framework called AlphaProof Code.

Formal verification, which involves using mathematics to check that software works as intended, has long been touted by its advocates as a holy grail of software security. It allows for verifying that a program contains no critical flaws. However, its large-scale use remains uncertain, whether regarding its widespread practical application or its effectiveness in securing large amounts of code.

Before the AI era, the task was extremely time-consuming: one had to describe the characteristics of the software to be proven, translate these descriptions into formulas, and then develop the proof. According to the source, agentic AI can significantly reduce the time needed for this last step, which made an analysis like the one conducted on Signal possible. This analysis remains limited, however. It did not, for example, allow for formally verifying the entire Signal mobile application, and the DeepMind researchers stress that their results represent only a step in a broader process.

Mixed Results on Older Versions

The analysis found no new vulnerabilities in the examined parts of the protocol. When the researchers applied AlphaProof Code to older versions of Signal containing long-known and fixed flaws, the tool correctly identified them and showed that the protocol could not be formally verified in those versions.

Swarat Chaudhuri, a research scientist at DeepMind, summarized the remaining path: “We are very excited about this result, but there is a long distance to go before we can scale this approach to all software in the world, and we want to get there.” The researchers are publishing the specification and proof developed for the Signal protocol analysis to facilitate other formal verification work.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x