PV3 ⊧ ERAN: ETH Robustness Analyzer for Neural Networksverifies safety properties of neural networks against input perturbations

An analyzer based on abstract interpretation for verification of MNIST, CIFAR10, and ACAS Xu based networks.

Application domain/field

• Neural networks
• Abstract interpretation
• Safety properties

Expected input

• Neural network
• Dataset
• Lots of parameters, depending on what analysis you want to run.

Format:

• Neural network: ONNX
• Dataset: Argument passed to ERAN. It has mnist, acasxu, cifar10

Expected output

Depends on the analysis that is run.

Internals

• Has its own ERAN language

Comments

License: Apache-2.0