DeformRS: Certifying input deformations with randomized smoothing
Deep neural networks are vulnerable to input deformations in the form of vector fields of pixel displacements and to other parameterized geometric deformations e.g. translations, rotations, etc. Current input deformation certification methods either 1. do not scale to deep networks on large input da...
| Main Authors: | , , , , |
|---|---|
| פורמט: | Conference item |
| שפה: | English |
| יצא לאור: |
Association for the Advancement of Artificial Intelligence
2022
|
| _version_ | 1826308194477015040 |
|---|---|
| author | Alfarra, M Bibi, A Khan, N Torr, P Ghanem, B |
| author_facet | Alfarra, M Bibi, A Khan, N Torr, P Ghanem, B |
| author_sort | Alfarra, M |
| collection | OXFORD |
| description | Deep neural networks are vulnerable to input deformations in the form of vector fields of pixel displacements and to other parameterized geometric deformations e.g. translations, rotations, etc. Current input deformation certification methods either 1. do not scale to deep networks on large input datasets, or 2. can only certify a specific class of deformations, e.g. only rotations. We reformulate certification in randomized smoothing setting for both general vector field and parameterized deformations and propose DeformRS-VF and DeformRS-Par, respectively. Our new formulation scales to large networks on large input datasets. For instance, DeformRS-Par certifies rich deformations, covering translations, rotations, scaling, affine deformations, and other visually aligned deformations such as ones parameterized by Discrete-Cosine-Transform basis. Extensive experiments on MNIST, CIFAR10, and ImageNet show competitive performance of DeformRS-Par achieving a certified accuracy of 39% against perturbed rotations in the set [-10°,10°] on ImageNet. |
| first_indexed | 2024-03-07T07:14:22Z |
| format | Conference item |
| id | oxford-uuid:cc0b9d8c-5acd-4ac8-9f56-d3066df20c1b |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-03-07T07:14:22Z |
| publishDate | 2022 |
| publisher | Association for the Advancement of Artificial Intelligence |
| record_format | dspace |
| spelling | oxford-uuid:cc0b9d8c-5acd-4ac8-9f56-d3066df20c1b2022-07-22T10:43:07ZDeformRS: Certifying input deformations with randomized smoothingConference itemhttp://purl.org/coar/resource_type/c_5794uuid:cc0b9d8c-5acd-4ac8-9f56-d3066df20c1bEnglishSymplectic ElementsAssociation for the Advancement of Artificial Intelligence2022Alfarra, MBibi, AKhan, NTorr, PGhanem, BDeep neural networks are vulnerable to input deformations in the form of vector fields of pixel displacements and to other parameterized geometric deformations e.g. translations, rotations, etc. Current input deformation certification methods either 1. do not scale to deep networks on large input datasets, or 2. can only certify a specific class of deformations, e.g. only rotations. We reformulate certification in randomized smoothing setting for both general vector field and parameterized deformations and propose DeformRS-VF and DeformRS-Par, respectively. Our new formulation scales to large networks on large input datasets. For instance, DeformRS-Par certifies rich deformations, covering translations, rotations, scaling, affine deformations, and other visually aligned deformations such as ones parameterized by Discrete-Cosine-Transform basis. Extensive experiments on MNIST, CIFAR10, and ImageNet show competitive performance of DeformRS-Par achieving a certified accuracy of 39% against perturbed rotations in the set [-10°,10°] on ImageNet. |
| spellingShingle | Alfarra, M Bibi, A Khan, N Torr, P Ghanem, B DeformRS: Certifying input deformations with randomized smoothing |
| title | DeformRS: Certifying input deformations with randomized smoothing |
| title_full | DeformRS: Certifying input deformations with randomized smoothing |
| title_fullStr | DeformRS: Certifying input deformations with randomized smoothing |
| title_full_unstemmed | DeformRS: Certifying input deformations with randomized smoothing |
| title_short | DeformRS: Certifying input deformations with randomized smoothing |
| title_sort | deformrs certifying input deformations with randomized smoothing |
| work_keys_str_mv | AT alfarram deformrscertifyinginputdeformationswithrandomizedsmoothing AT bibia deformrscertifyinginputdeformationswithrandomizedsmoothing AT khann deformrscertifyinginputdeformationswithrandomizedsmoothing AT torrp deformrscertifyinginputdeformationswithrandomizedsmoothing AT ghanemb deformrscertifyinginputdeformationswithrandomizedsmoothing |