Skip to content
main
Switch branches/tags
Go to file
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 

README.md

Fast and Complete: Complete Neural Network Verification with Rapid and Massively Parallel Incomplete Verifiers

We propose to use linear relaxation based perturbation analysis (LiRPA) to replace Linear Programming (LP) during the branch and bound (BaB) process in complete neural network verification. LiRPA based algorithms (e.g. CROWN and DeepPoly) can be efficiently implemented on typical machine learning accelerators such as GPUs and TPUs, and we exploit the automatic, highly efficient and differentiable implementation auto_LiRPA. We apply a fast gradient based bound tightening procedure combined with batch splits and the design of minimal usage of LP, which enable us to effectively use LiRPA on the accelerator hardware for the challenging complete NN verification problem. We significantly outperform LP-based approaches: on a single GPU, we demonstrate an order of magnitude speedup compared to existing LP-based approaches.

Please refer to our paper for more details:

Kaidi Xu*, Huan Zhang*, Shiqi Wang, Yihan Wang, Suman Jana, Xue Lin and Cho-Jui Hsieh, "Fast and Complete: Enabling Complete Neural Network Verification with Rapid and Massively Parallel Incomplete Verifiers", pdf (* Equal contribution)

Our implementation is based on the auto_LiRPA library. We are working on finalizing our code and will release full source code soon.

About

Code for paper "Fast and Complete: Enabling Complete Neural Network Verification with Rapid and Massively Parallel Incomplete Verifiers"

Resources

Releases

No releases published

Packages

No packages published