Adversarial purification with one-step guided diffusion model

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Diffusion models show promise for adversarial purification due to their generalization capabilities.
• Current diffusion models require multi-step sampling, leading to high time and resource consumption.
• Efficient defense mechanisms are crucial for robust machine learning models against adversarial attacks.

Purpose of the Study:

• To propose a novel One-Step Guided Diffusion Model (OSGD) for efficient adversarial purification.
• To accelerate the sampling speed of diffusion models for adversarial purification.
• To achieve superior purification results and enhance defense against adversarial examples.

Main Methods:

• Developed a One-Step Guided Diffusion Model (OSGD) incorporating a one-step denoising process.
• Implemented a guiding strategy using preliminary denoised images from a first round of denoising.
• Applied two rounds of one-step denoising, with the second round guided by the first-round output.

Main Results:

• OSGD significantly accelerates sampling speed compared to traditional multi-step diffusion models.
• The proposed guidance strategy effectively eliminates adversarial perturbations, improving purification results.
• Achieved state-of-the-art performance in accuracy and efficiency on Cifar10 and ImageNet datasets against PGD, AutoAttack, and BPDA+EOT attacks.

Conclusions:

• OSGD offers an efficient and effective solution for adversarial purification.
• The one-step guided denoising approach demonstrates the viability of preliminary denoised images as guidance signals.
• The method provides a strong defense against various sophisticated adversarial attack strategies.