Scalable Polyhedral Verification of Recurrent Neural Networks

We present a scalable and precise verifier for recurrent neural networks, called Prover. Prover relies on two novel ideas: (i) a method to compute a set of polyhedral abstractions for the non-convex and non-linear recurrent update functions by combining sampling, optimization, and Fermat’s theorem, and (ii) a gradient descent based algorithm for abstraction refinement guided by the certification problem that combines multiple abstractions for each neuron. Using Prover, we present the first study of certifying a non-trivial use case of recurrent neural networks, namely speech classification. To achieve this, we additionally develop custom abstractions for the non-linear speech preprocessing pipeline. Our evaluation shows that Prover successfully verifies several challenging recurrent models in computer vision and speech classification beyond the reach of prior work.

Please follow the `readme.md` located in the home directory. The system requires:

Gurobi optimizer licence - please refer to https://www.gurobi.com/academia/academic-program-and-licenses/
GPU available environment (not necessary but highly recommended)

Preliminary information for the VM:
1. List of essential files and directory
* /home/cav2021-paper187/readme.md: includes detailed instructions about how to prepare and run the artifact.
* /home/cav2021-paper187/CAV2021_paper_187.pdf: the accepted paper pdf.
* /home/cav2021-paper187/artifact/: the artifact codes and necessaries.
2. Recommended resource allocaiton
* Memory: 4GB
* Disk: 20GB because of the dataset
* GPU available environment
3. Credentials
* ID: cav2021-paper187
* PW: 2021-187-prover