A major research initiative developing sound, scalable formal verification tools and cryptographic proofs for deep neural network execution across distributed computing nodes.
How can deep neural networks deployed across untrusted distributed environments be verified for correctness, robustness, and mathematical safety with minimal computational overhead?
The widespread adoption of foundation models in mission-critical applications—from autonomous robotics to clinical diagnosis—demands mathematical guarantees rather than purely empirical testing. Project VERIPROVE tackles this challenge by creating a unified theoretical and algorithmic foundation for verifying complex neural models.
Developing tight numeric abstractions that propagate input bounds through non-linear activation functions without exponential state explosion.
Engineering lightweight cryptographic attestation proofs that prove a cloud computing instance executed a specific neural network without altering weights or precision.
Connecting neural verification routines directly to Lean 4 and Coq proof assistants, enabling end-to-end formal mathematical certification.
Scientific Data Vol. 3, pp. 160018(2016). DOI: 10.1038/sdata.2016.18
Defines the principles that digital research objects should be Findable, Accessible, Interoperable, and Reusable for people and machines.
Nature Human Behaviour Vol. 1, pp. 0021(2017). DOI: 10.1038/s41562-016-0021
Sets out practical measures to improve the reliability and efficiency of scientific research through stronger methods, transparency, and incentives.