Please use this identifier to cite or link to this item: https://hdl.handle.net/10356/161935
Title: Exploiting the relationship between Kendall’s rank correlation and cosine similarity for attribution protection
Authors: Wang, Fan
Kong, Adams Wai Kin 
Keywords: Computer Science - Learning
Computer Science - Artificial Intelligence
Issue Date: 2022
Source: Wang, F. & Kong, A. W. K. (2022). Exploiting the relationship between Kendall’s rank correlation and cosine similarity for attribution protection. Thirty-Sixth Conference on Neural Information Processing Systems (NeurIPS 2022).
Project: RG73/21 
Conference: Thirty-Sixth Conference on Neural Information Processing Systems (NeurIPS 2022)
Abstract: Model attributions are important in deep neural networks as they aid practitioners in understanding the models, but recent studies reveal that attributions can be easily perturbed by adding imperceptible noise to the input. The non-differentiable Kendall's rank correlation is a key performance index for attribution protection. In this paper, we first show that the expected Kendall's rank correlation is positively correlated to cosine similarity and then indicate that the direction of attribution is the key to attribution robustness. Based on these findings, we explore the vector space of attribution to explain the shortcomings of attribution defense methods using $\ell_p$ norm and propose integrated gradient regularizer (IGR), which maximizes the cosine similarity between natural and perturbed attributions. Our analysis further exposes that IGR encourages neurons with the same activation states for natural samples and the corresponding perturbed samples, which is shown to induce robustness to gradient-based attribution methods. Our experiments on different models and datasets confirm our analysis on attribution protection and demonstrate a decent improvement in adversarial robustness.
URI: https://hdl.handle.net/10356/161935
URL: https://nips.cc/
Schools: School of Computer Science and Engineering 
Interdisciplinary Graduate School (IGS) 
Research Centres: Rapid-Rich Object Search (ROSE) Lab 
Rights: © 2022 The Author(s). All rights reserved. This paper was published in the Proceedings of Thirty-Sixth Conference on Neural Information Processing Systems (NeurIPS 2022) and is made available with permission of The Author(s).
Fulltext Permission: open
Fulltext Availability: With Fulltext
Appears in Collections:IGS Conference Papers
SCSE Conference Papers

Files in This Item:
File Description SizeFormat 
nips22.pdf2.15 MBAdobe PDFThumbnail
View/Open

Page view(s)

146
Updated on Feb 24, 2024

Download(s) 50

58
Updated on Feb 24, 2024

Google ScholarTM

Check

Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.