Adversarial guided diffusion models for adversarial purification.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Diffusion models (DM) offer powerful adversarial purification (AP) for threat removal.
• Existing methods may compromise semantic information or preserve perturbations.
• Guided diffusion techniques struggle with distance measures that can retain perturbations.

Purpose of the Study:

• To develop a novel adversarial guided diffusion model for robust adversarial purification.
• To enhance the preservation of semantic information during perturbation removal.
• To improve the overall robustness of diffusion model-based adversarial purification.

Main Methods:

• Proposed an adversarial guided diffusion model with a novel adversarial guidance mechanism.
• Utilized an auxiliary neural network trained adversarially for guidance.
• Employed distance measures in latent representations, not pixel-level values, for guidance.

Main Results:

• Demonstrated effectiveness in simultaneously maintaining semantic information and removing adversarial perturbations.
• Achieved significant enhancements in the robustness of existing DM-based AP methods.
• Reported an average robust accuracy improvement of up to 7.30% on CIFAR-10.

Conclusions:

• The proposed adversarial guidance effectively balances semantic preservation and perturbation removal.
• The method significantly boosts the performance of diffusion model-based adversarial purification.
• This approach offers a promising direction for enhancing the security of deep learning models against adversarial attacks.