=Paper=
{{Paper
|id=Vol-2908/invited1
|storemode=property
|title=Invited Talk: Using SMT and Abstraction-Refinement for Neural Network Verification
|pdfUrl=https://ceur-ws.org/Vol-2908/invited1.pdf
|volume=Vol-2908
|authors=Guy Katz
|dblpUrl=https://dblp.org/rec/conf/smt/Katz21
}}
==Invited Talk: Using SMT and Abstraction-Refinement for Neural Network Verification==
<pdf width="1500px">https://ceur-ws.org/Vol-2908/invited1.pdf</pdf>
<pre>
Invited Talk: Using SMT and Abstraction-Refinement
for Neural Network Verification
Guy Katz1 , Yizhak Yisrael Elboher2
1
    The Hebrew University of Jerusalem, Jerusalem, Israel
2
    The Hebrew University of Jerusalem, Jerusalem, Israel


                                         Abstract
                                         Deep neural networks are increasingly being used as controllers for safety-critical systems. Because
                                         neural networks are opaque, certifying their correctness is a significant challenge. To address this issue,
                                         several neural network verification approaches have recently been proposed, many of them based on
                                         SMT solving. However, these approaches afford limited scalability, and applying them to large networks
                                         can be challenging. In this talk we will discuss a framework that can enhance neural network verification
                                         techniques by using over-approximation to reduce the size of the network — thus making it more
                                         amenable to verification. This approximation is performed such that if the property holds for the smaller
                                         (abstract) network, it holds for the original as well. The over-approximation may be too coarse, in which
                                         case the underlying verification tool might return a spurious counterexample. Under such conditions, we
                                         can perform counterexample-guided refinement to adjust the approximation, and then repeat the process.
                                         This approach is orthogonal to, and can be integrated with, many existing verification techniques. For
                                         evaluation purposes, we integrate it with the recently proposed Marabou framework, and observe a
                                         significant improvement in Marabou’s performance. Our experiments demonstrate the great potential of
                                         abstraction-refinement for verifying larger neural networks.




SMT’21: 19th International Workshop on Satisfiability Modulo Theories, July 18–19, 2021, Online
" g.katz@mail.huji.ac.il (G. Katz); yizhak.elboher@mail.huji.ac.il (Y. Y. Elboher)
                                       © 2021 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
    CEUR
    Workshop
    Proceedings
                  http://ceur-ws.org
                  ISSN 1613-0073       CEUR Workshop Proceedings (CEUR-WS.org)



                                                                                                           1

</pre>