Image classification adversarial attack with improved resizing transformation and ensemble models.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Convolutional neural networks (CNNs) are powerful in computer vision but vulnerable to adversarial examples.
• Adversarial examples, imperceptible input perturbations, highlight CNN vulnerabilities and are used to assess network robustness.
• Black-box attacks have lower success rates and limited transferability compared to white-box attacks.

Purpose of the Study:

• To improve the success rate and transferability of adversarial examples in black-box attacks.
• To introduce a novel model augmentation technique for generating more robust adversarial examples.
• To enhance the security evaluation of neural networks against sophisticated adversarial attacks.

Main Methods:

• Proposed a resizing invariance method for model augmentation, inspired by data augmentation techniques.
• Utilized improved resizing transformations to enhance model augmentation capabilities.
• Employed ensemble models to generate adversarial examples with increased transferability.

Main Results:

• The proposed resizing invariance method demonstrated superior performance compared to baseline methods.
• Achieved significant improvements in black-box attack success rates on both normal and defense models.
• Validated the effectiveness of ensemble models in generating more transferable adversarial examples.

Conclusions:

• The resizing invariance method is an effective approach for model augmentation in the context of adversarial attacks.
• The proposed techniques enhance the transferability of adversarial examples, posing a greater challenge to model security.
• This research contributes to a better understanding and evaluation of neural network robustness against adversarial perturbations.