Image Adversarial Example Generation Method Based on Adaptive Parameter Adjustable Differential Evolution

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Adversarial attacks on neural networks are a growing concern.
• Generating effective adversarial examples with minimal changes is challenging.

Purpose of the Study:

• To propose a novel adversarial example generation method.
• To achieve high attack success rates with minimal pixel modifications.

Main Methods:

• Adaptive parameter adjustable differential evolution algorithm.
• Dynamic adjustment of control parameters and operation strategies.
• Searching for optimal pixel perturbations.

Main Results:

• Generated adversarial examples with high success rates and few pixel modifications.
• Outperformed conventional differential evolution-based One Pixel Attack.
• Required significantly less perturbation than global/local attacks.
• Demonstrated increased resistance to perception and detection.

Conclusions:

• The proposed adaptive differential evolution method is effective for generating adversarial examples.
• The method offers a superior balance between attack success and perturbation magnitude.
• It presents a more robust adversarial attack strategy for image recognition models.