적대적 공격 방어를 위한 특이값 기반 무작위 이미지 재구성 기법
Singular Value-Based Randomized Image Reconstruction Method for Adversarial Attack Defense
2
조회
0
다운로드
1
인용
0
참고문헌
키워드
초록
인공지능이 발달함에 따라 인공지능의 오판을 야기하는 적대적 공격 기법도 발전하였다. 이를 방어하기 위해 주로 신경망의 강건성을 증가시키는 데 초점을 맞춰 연구가 이루어졌다. 하지만 기존 방어 기술들은 다양한 공격 방법에 대해 강건하지 않아 새로운 유형의 공격에는 취약하다. 또한 방어율을 향상하는 동시에 모델의 정확도를 유지하기는 매우 어렵다. 본 연구에서는 다양한 공격 기법에 대해 균일하게 방어하면서 신경망의 정확도도 유지할 수 있는 새로운 신경망 훈련 및 추론 기법을 제안한다. 본 연구에서 제안하는 알고리즘은 특이값 분해(SVD)를 사용하여 무작위화 기반의 이미지 재구성을 수행하고, 이를 통해 입력 데이터를 공격자가 예측할 수 없게 만들어 적대적 공격을 방어한다. CIFAR-10 데이터 세트에서 제안하는 방법을 실험하여 간단하면서도 균일한 방어가 가능한 방어 기술임을 확인하였고, FGSM 공격에 대해서는 최종 방어율 61%를 기록하여 성능을 입증하였다.
Abstract
As AI has evolved, so have adversarial attacks that cause AI to misjudge. To defend against this, research has focused primarily on increasing the robustness of neural networks. However, existing defense techniques are not robust against a variety of attack methods, leaving them vulnerable to new types of attacks. It is also very difficult to improve the defense rate while maintaining the network accuracy. In this work, we propose a novel neural network training and inference technique that can uniformly defend against various attack methods while maintaining the accuracy of the network. The algorithm we propose in this work performs randomization-based image reconstruction using singular value decomposition to defend against adversarial attacks by making the input data unpredictable to the attacker. We experimented with the method on the CIFAR-10 dataset and found it to be a simple and uniform defense technique. The method's performance is demonstrated by achieving a final defense rate for 61% against FGSM attacks.
참고문헌 (19)
- I. J. Goodfellow, "Explaining and harnessing adversarial examples" (2014)
- D. Amodei, "Concrete problems in ai safety" (2016)
- K. Eykholt, "Robust physical-world attacks on deep learning visual classification", Proceedings of the IEEE conference on computer vision and pattern recognition, pp.1625~1634 (2018)
- N. Kanagaraj, "Deep learning using computer vision in self driving cars for lane and traffic sign detection", International Journal of System Assurance Engineering and Management, 12(6), pp.1011~1025 (2021)
- D. Shen, "Deep learning in medical image analysis", Annual review of biomedical engineering, 19, pp.221~248 (2017)
- A. Kurakin, "Adversarial machine learning at scale" (2016)
- U. Shaham, "Understanding adversarial training : Increasing local stability of supervised models through robust optimization", Neurocomputing, 307, pp.195~204 (2018)
- C. Szegedy, "Intriguing properties of neural networks" (2013)
- N. Papernot, "Distillation as a defense to adversarial perturbations against deep neural networks", 2016 IEEE Symposium on Security and Privacy (SP) (2016)
- N. Papernot, "Extending defensive distillation" (2017)
- F. Tramr, "Ensemble adversarial training: Attacks and defenses", International Conference on Learning Representations, 1 (2018)
- N. Carlini, "Towards evaluating the robustness of neural networks", 2017ieee symposium on security and privacy (sp), pp.39~57 (2017), Ieee
- W. Xu, "Feature squeezing:Detecting adversarial examples in deep neural networks" (2017)
- S.-M. Moosavi-Dezfooli, "Deepfool: a simple and accurate method to fool deep neural networks", Proceedings of the IEEE conference on computer vision and pattern recognition, pp.2574~2582 (2016)
- A. Madry, "Towards deep learning models resistant to adversarial attacks" (2017)
- A. Kurakin, "Artificial intelligence safety and security", pp.2018, Chapman and Hall
- K. Simonyan, "Very deep convolutional networks for large-scale image recognition" (2014)
- J. Rauber, "Foolbox native: Fast adversarial attacks to benchmark the robustness of machine learning models in pytorch, tensorflow, and jax", Journal of Open Source Software, 5(53), pp.2607~ (2020)
- Dong, Minjing, "Adversarial Robustness via Random Projection Filters", Proceedings of the IEEE, 2023, ~