적대적 공격 및 방어 기술의 성능 향상을 위한 역방향 적대적 데이터 생성 연구
Reverse-Update Adversarial Data for Enhancing Adversarial Attack and Adversarial Training Performance
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
인공지능 기술의 오작동을 유도하는 적대적 공격은 다양한 도메인과 모델에 적용 가능하며, 성능이 높은 SOTA (State-of-the-Art) 모델의 성능도 손쉽게 저해 시킬 수 있다. 이에 대처하기 위해 적대적 방어 기술들이 개발되고 있지만, 명확한 한계점으로 인해 활용이 제한된다. 그 결과, 특정 분야에서 인공지능 기술의 도입 뿐만 아니라 고도화 연구도 지체되고 있다. 해당 문제를 해결하기 위해, 본 논문에서는 적대적 공격의 손실 함수 업데이트 방향의 부호를 바꿔 새로운 개념의 적대적 데이터를 소개한다. 본 연구에서 소개한 역방향 적대적 데이터를 데이터 오염 및 적대적 훈련 환경에 적용하여 실험을 진행한 결과, 모델의 성능을 최대 72% 낮추고 9개 환경 중 6개 환경에서 강건성 향상에 가장 효율적임을 입증했다. 결과적으로, 제시한 적대적 데이터는 적대적 공격과 방어 기술의 연구 확장을 유도할 수 있으며, 더 나아가 방어 기술 개발의 고도화를 촉진할 수 있어 AI의 안전한 도입에 기여한다.
Abstract
Adversarial attacks, which induce malfunctions in AI technologies, can be applied to various domains and models, easily compromising SOTA (State-of-the-Art) models. Although adversarial defense techniques have been developed to counter these attacks, their applicability is limited due to constraints. Consequently, not only is the adoption of AI technology delayed, but also advanced research is restricted. To address this issue, this paper introduces a novel concept of adversarial data by reversing the sign of the loss function update in adversarial attacks. Experiments were conducted by applying the reverse-update adversarial data to data poisoning and adversarial training environments, showing that it can reduce the model’s performance up to 72% and is most effective in enhancing robustness in 6 out of 9 environments. Ultimately, the proposed data can stimulate extensive research on adversarial attacks and defenses, further promoting the advancement of defense technology and contributing to the safe adoption of AI.
참고문헌 (18)
- Sun Lu, "A survey of practical adversarial example attacks", Cybersecurity, 1(1), pp.9~ (2018)
- S.A. Rebuffi, "icarl: Incremental classifier and representation learning", Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pp.2001~2010 (2017)
- Rahaf Aljundi, "Gradient based sample selection for online continual learning", Advances in neural information processing systems, pp.10~ (2019)
- Mallya, Arun, "Packnet: Adding multiple tasks to a single network by iterative pruning", Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pp.7765~7773 (2018)
- Purushwalkam, Senthil, "The challenges of continuous self-supervised learning", European Conference on Computer Vision, pp.702~721 (2022)
- Matteo Boschini, "Transfer without forgetting", European Conference on Computer Vision, pp.692~709 (2022)
- S.I. Mirzadeh, "Understanding the role of training regimes in continual learning", Advances in Neural Information Processing Systems, 33, pp.7308~7320 (2020)
- Shipeng Wang, "Training networks in null space of feature covariance for continual learning", Proceedings of the IEEE, 2021, 184~193
- Li, Zhizhong, "Learning without forgetting", IEEE transactions on pattern analysis and machine intelligence, 40(12), pp.2935~2947 (2017)
- Lin, Guoliang, "Towards better plasticity-stability trade-off in incremental learning: A simple linear connector", Proceedings of the IEEE, 2022, 89~98
- Kurakin, Alexey, "Adversarial examples in the physical world", Artificial intelligence safety and security, pp.99~112 (2018)
- Carlini, Nicholas, "Towards evaluating the robustness of neural networks", 2017 ieee symposium on security and privacy, pp.39~57 (2017)
- Nicolas Papernot, "Distillation as a defense to adversarial perturbations against deep neural networks", 2016 IEEE symposium on security and privacy, pp.582~597 (2016)