Exploring Targeted Misclassification Attacks: Leveraging GradCam for Image Manipulation and Label Misclassification

This research paper conducts a comprehensive investigation into the susceptibility of deep learning models to adversarial attacks, specifically focusing on targeted misclassification attacks and their implications for applications utilizing these models. The paper initially provides an overview of deep learning models, highlighting their significance across various domains, and then delves into the concept of adversarial attacks, emphasizing their ability to manipulate deep learning models and compromise their reliability. The study explores targeted misclassification attacks in-depth, discussing their motivations and potential consequences for deep learning-based applications. To assess the impact of targeted misclassification attacks, the paper employs the GradCam method, which enables the modification of images based on the GradCam of the desired target class. By adopting this approach, the study aims to reveal the vulnerability of deep learning models to targeted misclassification attacks, offering insights into potential defense mechanisms and underscoring the importance of safeguarding deep learning-based applications against evolving adversarial threats. The experimental results demonstrate the effectiveness of the proposed approach, achieving a favorable average fooling ratio of ۰.۷۰ and an average rate of ۰.۳۶ for adversarial confidence drop in generating deceptive adversarial samples.

Keywords:

deep learning models , adversarial attacks , targeted misclassification , GradCam , image modification , model misclassification , defense mechanisms