Robust image classification against adversarial attacks using elastic similarity measures between edge count...

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep neural networks (DNNs) excel at pattern recognition but are vulnerable to adversarial attacks.
• Adversarial images are subtly manipulated to fool DNNs without noticeable visual changes.
• Existing defenses struggle against advanced adversarial attack strategies.

Purpose of the Study:

• To develop a novel method for improving the robustness of DNNs in image classification against adversarial attacks.
• To leverage the human visual system's sensitivity to geometric anomalies over color variations for attack detection.
• To introduce a discrimination module that analyzes color gradient features to identify manipulated images.

Main Methods:

• Extracting color gradient features from input images at multiple sensitivity levels.
• Utilizing a deep neural classifier for image categorization.
• Employing a discrimination model with time series analysis to detect adversarial manipulations based on color gradients.

Main Results:

• The proposed discrimination module significantly enhances DNN robustness against adversarial attacks.
• Adversarial attack success rates were drastically reduced, particularly for whole-image attacks.
• The method demonstrated effectiveness against state-of-the-art adversarial attack techniques.

Conclusions:

• The color gradient analysis method effectively improves DNN robustness in image classification.
• This approach offers a promising defense against visually imperceptible adversarial manipulations.
• Future work will focus on generalizing the detection accuracy for broader attack strategies.