Earticle

초록

영어

Deep neural networks (DNNs) are widely deployed across safety-critical applications but remain vulnerable to adversarial attacks. This vulnerability is especially severe for compact models, which lack sufficient capacity to learn robust decision boundaries. While adversarial distillation (AD) offers a promising solution by transferring robustness from a large, robust teacher to a lightweight student, a fundamental limitation persists: as training progresses, the student generates increasingly strong adversarial examples that diverge from the teacher’s original adversarial domain. Since the teacher is typically fixed in conventional AD frameworks, its predictions become progressively less reliable on student-generated adversarial inputs, resulting in degraded supervision and limited robustness gains. To overcome this limitation, we propose ProxyAD, which equips the frozen teacher with a lightweight ProxyNet to dynamically align supervision with the student’s evolving adversarial inputs. ProxyNet is trained on student-generated adversarial examples using the teacher’s prediction on the corresponding unperturbed input as supervision, while a studentalignment regularizer prevents overconfident proxy soft labels on adversarial inputs by pulling them toward the student’s distribution so that targets remain calibrated and learnable during AD. This resulting supervision improves the student’s robustness with minimal additional parameters and modest training cost. Extensive experiments on CIFAR10/100 and Tiny-ImageNet show that ProxyAD consistently outperforms existing AD methods in both clean and adversarial accuracy under various attack scenarios.

목차

Abstract
1. Introduction
2. Related Work
3. Preliminary
3. Adversarial Distillation via Proxy Teacher Adaptation (ProxyAD)
3.1 Motivation and Overview
3.2 Teacher Adaptation via ProxyNet
3.3 Student Training Objective in AD framework
3.4 Overall Training Procedure
3.4 Overall Training Procedure
4. Results
4.1 Experimental Setup
4.2 Main Results
4.3 Self-Distillation Results
5. Conclusion
References

키워드

저자

• Hyejin Park [ Ph.D. Student, Department of Computer Science and Engineering, Ewha Womans University, Seoul, Korea/Visiting Professor, Department of Media Software, Sungkyul University, Anyang, Korea ] Corresponding Author

참고문헌

발행기관

간행물

표지보기 관심저널 등록

• 간행물명

  International Journal of Internet, Broadcasting and Communication

• 간기

  계간

• pISSN

  2288-4920

• eISSN

  2288-4939

• 수록기간

  2009~2025

• 십진분류

  KDC 326 DDC 380