Imperceptible adversarial perturbations via temporal summation in image sequences = 시간적 합산을 활용한 이미지 시퀀스 적대적 섭동의 비가시성 향상

다국어 초록 (Multilingual Abstract)

In this paper, we present a new methodology for improving the imperceptibility of adversarial perturbations for image sequences. We introduce temporal summation loss, which is designed to minimize the temporal summation of adversarial perturbations, leveraging an effect of the human eye: integrating brightness over a certain period. We integrate this temporal summation loss with the existing projected gradient descent (PGD) methodology to introduce PGD-TL. Through experiments, PGD-TL was quantitatively proven, via temporal summation loss, to significantly enhance the imperceptibility of adversarial perturbations while maintaining a similar attack success rate to existing methodologies. Furthermore, a subjective test was used to analyze various perspectives, showing that our method improves imperceptibility compared to existing methods.

국문 초록 (Abstract)

해당 논문에서는 이미지 시퀀스에 대한 적대적 섭동의 비가시성을 향상시키기 위한 새로운 방법론을 제시한다. 특정 기간 동안 눈에 들어온 빛을 통합하는 사람의 인지적 특성을 이용하기 위해서 적대적 섭동의 시간적 합산을 최소화하도록 설계한 temporal summation loss를 제안한다. temporal summation loss와 기존의 적대적 공격 방법론인 projected gradient descent(PGD)을 결합하여 새로운 공격 방법론 PGD-TL을 소개한다. PGD-TL은 정량적 및 주관적인 실험을 통해 기존 공격 방법론과 동일한 공격 성공률을 유지하면서 적대적 섭동의 비가시성을 크게 향상시켰음을 보인다.

목차 (Table of Contents)

• Contents 1
• List of tables 2
• List of figures 3
• 1 Introduction 4
• 2 Background 6
• 2.1 Imperceptibility of adversarial examples 6
• 2.2 Temporal responsibility 7
• 3 Proposed method 8
• 3.1 Preliminaries 9
• 3.2 PGD-TL 10
• 3.3 Extensions of PGD-TL 16
• 4 Experiments 17
• 4.1 Experimental Setup 17
• 4.2 Quantitative evaluation 18
• 4.3 Qualitative evaluation 22
• 4.4 Analysis of PGD-TL extensions 27
• 4.5 Attack defense and limitation 29
• 5 Conclusion 30
• References 31
• 초 록 36
• Abstract 37