PV1 ⊧ SyReNN: Symbolic Representation for Neural Networkscomputes a symbolic representation of a neural network as linear functions

Application domain/field

• Deep neural networks (DNNs)
• Adversarial attacks
• Symbolic representation

Type of tool

Library for analyzing DNNs

Expected input

• Piecewise-linear DNN (Deep Neural Network)
• X: union of one- or two-dimensional polytopes

Format:

• DNN: ERAN or ONNX
• X: list of its vertices in counter-clockwise order

Expected output

Symbolic representation $\widehat{f_{\upharpoonright X}}$

Internals

Computes a symbolic representation that decomposes the behavior of a DNN into finitely-many linear functions. A DNN is piecewise-linear if its input domain can be precisely partitioned into finitely-many linear functions.