Adv-BDPM: Adversarial attack based on Boundary Diffusion Probability Model.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks (DNNs) exhibit remarkable performance but suffer from vulnerabilities to adversarial examples.
• Existing methods for generating adversarial perturbations often lack diversity and precision.
• Robustness of DNNs against adversarial attacks is a critical research area.

Purpose of the Study:

• To introduce a novel technique, Adversarial Boundary Diffusion Probability Modeling (Adv-BDPM), to address limitations in adversarial perturbation generation.
• To enhance the precision and diversity of adversarial attacks while improving DNN robustness.

Main Methods:

• Developed a boundary diffusion probability model by integrating denoising diffusion probability models with boundary loss.
• Generated boundary perturbations tailored to specific neural networks.
• Proposed a decision boundary search algorithm using iterative boundary and orthogonal perturbations to create adversarial samples.

Main Results:

• Adv-BDPM demonstrated a superior attack success rate and perturbation precision in black-box attacks on ImageNet.
• Experiments on CIFAR-10 and CIFAR-100 showed Adv-BDPM achieved higher attack success rates and improved attack diversity.
• The method effectively defended against adversarial training with reduced computational time.

Conclusions:

• Adv-BDPM offers a powerful new approach for generating adversarial examples, enhancing both attack effectiveness and network defense.
• The technique improves the precision and diversity of adversarial perturbations, contributing to more robust deep learning models.
• Adv-BDPM presents a computationally efficient solution for enhancing DNN security against adversarial threats.