A statistical approach to assessing neural network robustness
We present a new approach to assessing the robustness of neural networks based on estimating the proportion of inputs for which a property is violated. Specifically, we estimate the probability of the event that the property is violated under an input model. Our approach critically varies from the f...
Main Authors: | , , , |
---|---|
Format: | Conference item |
Published: |
International Conferences on Learning Representations
2019
|
_version_ | 1797083014511984640 |
---|---|
author | Webb, S Rainforth, T Teh, Y Mudigonda, P |
author_facet | Webb, S Rainforth, T Teh, Y Mudigonda, P |
author_sort | Webb, S |
collection | OXFORD |
description | We present a new approach to assessing the robustness of neural networks based on estimating the proportion of inputs for which a property is violated. Specifically, we estimate the probability of the event that the property is violated under an input model. Our approach critically varies from the formal verification framework in that when the property can be violated, it provides an informative notion of how robust the network is, rather than just the conventional assertion that the network is not verifiable. Furthermore, it provides an ability to scale to larger networks than formal verification approaches. Though the framework still provides a formal guarantee of satisfiability whenever it successfully finds one or more violations, these advantages do come at the cost of only providing a statistical estimate of unsatisfiability whenever no violation is found. Key to the practical success of our approach is an adaptation of multi-level splitting, a Monte Carlo approach for estimating the probability of rare events, to our statistical robustness framework. We demonstrate that our approach is able to emulate formal verification procedures on benchmark problems, while scaling to larger networks and providing reliable additional information in the form of accurate estimates of the violation probability. |
first_indexed | 2024-03-07T01:36:06Z |
format | Conference item |
id | oxford-uuid:953c3599-4cef-4a8a-a2a8-546877ea5597 |
institution | University of Oxford |
last_indexed | 2024-03-07T01:36:06Z |
publishDate | 2019 |
publisher | International Conferences on Learning Representations |
record_format | dspace |
spelling | oxford-uuid:953c3599-4cef-4a8a-a2a8-546877ea55972022-03-26T23:44:49ZA statistical approach to assessing neural network robustnessConference itemhttp://purl.org/coar/resource_type/c_5794uuid:953c3599-4cef-4a8a-a2a8-546877ea5597Symplectic Elements at OxfordInternational Conferences on Learning Representations2019Webb, SRainforth, TTeh, YMudigonda, PWe present a new approach to assessing the robustness of neural networks based on estimating the proportion of inputs for which a property is violated. Specifically, we estimate the probability of the event that the property is violated under an input model. Our approach critically varies from the formal verification framework in that when the property can be violated, it provides an informative notion of how robust the network is, rather than just the conventional assertion that the network is not verifiable. Furthermore, it provides an ability to scale to larger networks than formal verification approaches. Though the framework still provides a formal guarantee of satisfiability whenever it successfully finds one or more violations, these advantages do come at the cost of only providing a statistical estimate of unsatisfiability whenever no violation is found. Key to the practical success of our approach is an adaptation of multi-level splitting, a Monte Carlo approach for estimating the probability of rare events, to our statistical robustness framework. We demonstrate that our approach is able to emulate formal verification procedures on benchmark problems, while scaling to larger networks and providing reliable additional information in the form of accurate estimates of the violation probability. |
spellingShingle | Webb, S Rainforth, T Teh, Y Mudigonda, P A statistical approach to assessing neural network robustness |
title | A statistical approach to assessing neural network robustness |
title_full | A statistical approach to assessing neural network robustness |
title_fullStr | A statistical approach to assessing neural network robustness |
title_full_unstemmed | A statistical approach to assessing neural network robustness |
title_short | A statistical approach to assessing neural network robustness |
title_sort | statistical approach to assessing neural network robustness |
work_keys_str_mv | AT webbs astatisticalapproachtoassessingneuralnetworkrobustness AT rainfortht astatisticalapproachtoassessingneuralnetworkrobustness AT tehy astatisticalapproachtoassessingneuralnetworkrobustness AT mudigondap astatisticalapproachtoassessingneuralnetworkrobustness AT webbs statisticalapproachtoassessingneuralnetworkrobustness AT rainfortht statisticalapproachtoassessingneuralnetworkrobustness AT tehy statisticalapproachtoassessingneuralnetworkrobustness AT mudigondap statisticalapproachtoassessingneuralnetworkrobustness |