SCAAT: Improving Neural Network Interpretability via Saliency Constrained Adaptive Adversarial Training
Rui Xu (Peking University), Wenkang Qin (Peking University), Peixiang Huang (Peking University), Hao Wang (National Institutes for Food and Drug Control), Lin Luo (Peking University)*
The 34th British Machine Vision Conference
Abstract
Deep Neural Networks (DNNs) are expected to provide explanation for users to understand their black-box predictions. Saliency map is a common form of explanation illustrating the heatmap of feature attributions, but it suffers from noise in distinguishing important features. In this paper, we propose a model-agnostic learning method called Saliency Constrained Adaptive Adversarial Training (SCAAT) to improve the quality of DNN interpretability. By constructing adversarial samples under the guidance of saliency map, SCAAT effectively eliminates most noise and makes saliency maps sparser and more faithful without any modification to the model architecture. We apply SCAAT to multiple DNNs and evaluate the quality of the generated saliency maps on various natural and pathological image datasets. Evaluation on different domains and metrics show that SCAAT significantly improves the interpretability of DNNs by providing more faithful saliency maps and barely sacrifices their predictive power.Citation
@inproceedings{Xu_2023_BMVC,
author = {Rui Xu and Wenkang Qin and Peixiang Huang and Hao Wang and Lin Luo},
title = {SCAAT: Improving Neural Network Interpretability via Saliency Constrained Adaptive Adversarial Training},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0828.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by