Diffusion Models for Imperceptible and Transferable Adversarial Attack.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing adversarial attacks often create perceptible perturbations in image RGB space.
• Unrestricted attacks lack transferability to black-box models.
• Visual imperceptibility and transferability are key challenges in adversarial attacks.

Purpose of the Study:

• To propose a novel adversarial attack method, DiffAttack, that achieves both visual imperceptibility and transferability.
• To leverage the generative and discriminative capabilities of diffusion models for crafting adversarial examples.
• To address the limitations of existing adversarial attack strategies.

Main Methods:

• Crafting perturbations in the latent space of diffusion models instead of pixel space.
• Utilizing content-preserving structures to generate human-insensitive perturbations with semantic clues.
• Deceiving diffusion models by distracting their attention to improve transferability.

Main Results:

• DiffAttack generates imperceptible adversarial examples that are transferable across diverse model architectures (CNNs, Transformers, MLPs).
• The proposed method demonstrates superiority over iterative, GAN-based, and ensemble attacks.
• Experiments were conducted on ImageNet, CUB-200, and Standford Cars datasets, validating effectiveness against various defenses.

Conclusions:

• DiffAttack is the first method to integrate diffusion models into adversarial attacks, offering a new direction.
• The approach successfully balances imperceptibility and transferability, outperforming prior methods.
• Future research can explore further advancements in diffusion-based adversarial attacks.