PV1 ⊧ DLV: Deep Learning Verificationverifies that a neural network has no misclassifications

Framework for automated verification of safety of classification decisions made by feed-forward deep neural networks.

Application domain/field

• Deep neural networks
• Safety verification
• Image classification
• Adversarial perturbations
• Classifiers

Expected input

• Dataset
• Neural network
• ?

Format:

Everything seems to be defined within a python file.

Expected output

The network is verified (i.e. no misclassifications were found for all layers) or it is falsified (i.e. the tool found an adversarial example )

Internals

Uses Z3, Keras (neural network library), Theano (deep learning package).

Comments

License: GNU General Public License v3.0