An Intermediate-Level Attack Framework on the Basis of Linear Regression.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Adversarial examples in machine learning are crucial for understanding model robustness.
• Improving the transferability of adversarial examples is a key challenge in developing effective attacks.
• Previous work introduced intermediate-level attacks to enhance adversarial example transferability.

Purpose of the Study:

• To extend previous research on intermediate-level attacks for improved adversarial example transferability.
• To propose and analyze a framework establishing a direct linear mapping from intermediate-level feature discrepancies to prediction loss.
• To achieve new state-of-the-art results in transfer-based adversarial attacks.

Main Methods:

• Developed a framework utilizing linear regression models to map feature discrepancies to prediction loss.
• Investigated the correlation between the magnitude of intermediate-level adversarial discrepancies and attack transferability.
• Employed multiple runs of baseline attacks with random initialization to boost performance.

Main Results:

• Demonstrated that various linear regression models can effectively establish the desired mapping.
• Confirmed a positive correlation between the magnitude of intermediate-level adversarial discrepancies and attack transferability.
• Achieved new state-of-the-art performance on transfer-based infinity (l∞) and L2 adversarial attacks.

Conclusions:

• The proposed framework significantly enhances the transferability of adversarial examples.
• Linear mapping from feature discrepancies offers a promising direction for robust adversarial attack development.
• The findings pave the way for more effective and transferable adversarial attacks in computer vision.