Earticle

초록

영어

This research aims to identify the types of adversarial attacks on AI image recognition models applicable to real-world security threat scenarios, evaluate their risk level and detection difficulty, and contribute to the development of robust and reliable defense measures. We identified various attack types and risk levels against deep learning-based image recognition models, and presented the characteristics and limitations of Projected Gradient Descent-based attacks. Based on this analysis, we designed experiments for different Projected Gradient Descent variants, compared their performance, and quantitatively evaluated their real-world attack probability and detection evasion. The experimental results show that the Early Stop- Projected Gradient Descent model has the highest attack performance compared to other Projected Gradient Descent-based attack models, which is a good trade-off between attack strength control and detection avoidance. We analyzed the risk by integrating attack type, medium, risk level, and detection difficulty, and proposed a unified view that enables a structural understanding of the attack-detection interaction, beyond the limitations of previous studies that are limited to studying individual techniques. This research contributes to the field of AI image recognition security by integrating attack experimentation, code improvement, risk identification, and detectability analysis.

목차

Abstract
1. Introduction
2. Related Research
3. AI Image Recognition Attacks
3.1 AI Image Recognition Attacks
3.2 Risk level of AI image recognition attacks and detection
3.3 Differences from previous studies
3.4 Features of Projected Gradient Descent Noise Attack
4. Conclusions
Acknowledgement
References

키워드

저자

• Jin-keun Hong [ Professor, Division of Advanced IT, Baekseok University, Korea ] Corresponding Author

참고문헌

발행기관

간행물

표지보기 관심저널 등록

• 간행물명

  The International Journal of Advanced Smart Convergence

• 간기

  계간

• pISSN

  2288-2847

• eISSN

  2288-2855

• 수록기간

  2012~2025

• 십진분류

  KDC 326 DDC 380