Document Type

Article

Source of Publication

IEEE Access

Publication Date

9-11-2024

Abstract

Deep neural networks (DNNs), while powerful, often suffer from a lack of interpretability and vulnerability to adversarial attacks. Concept bottleneck models (CBMs), which incorporate intermediate high-level concepts into the model architecture, promise enhanced interpretability. This study delves into the robustness of Concept Bottleneck Models (CBMs) against adversarial attacks, comparing their original and adversarial performance with standard Convolutional Neural Networks (CNNs). The premise is that CBMs prioritize conceptual integrity and data compression, enabling them to maintain high performance under adversarial conditions by filtering out non-essential variations in input data. Our extensive evaluations across different datasets and adversarial attacks confirm that CBMs not only maintain higher accuracy but also show improved defense capabilities against a range of adversarial attacks compared to traditional models. Our findings indicate that CBMs, particularly those trained sequentially, inherently exhibit higher robustness against adversarial attacks than their standard CNN counterparts. Additionally, we explore the effects of increasing conceptual complexity and the application of adversarial training techniques. While adversarial training generally boosts robustness, the increment varies between CBMs and CNNs, highlighting the role of training strategies in achieving adversarial resilience.

ISSN

2169-3536

Publisher

Institute of Electrical and Electronics Engineers (IEEE)

First Page

131323

Last Page

131335

Disciplines

Computer Sciences

Keywords

Adversarial attacks, Concept Bottleneck models, Interpretable models, Robustness

Scopus ID

85204159105

Indexed in Scopus

yes

Open Access

yes

Open Access Type

Gold: This publication is openly available in an open access journal/series

Download

Share

COinS