Diffusion Models as Strong Adversaries.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Diffusion models excel at generating high-quality images.
• Their adversarial capabilities, especially in no-box attack scenarios, remain underexplored.
• Existing no-box attacks often require access to training data, limiting real-world applicability.

Purpose of the Study:

• To investigate the adversarial potential of diffusion models.
• To propose a novel no-box adversarial attack method using only diffusion model-generated data.
• To evaluate the attack's effectiveness on the ImageNet dataset.

Main Methods:

• Generated a synthetic dataset using diffusion models to train a substitute model.
• Fine-tuned the substitute model using a classification diffusion model, incorporating model uncertainty and noise augmentation.
• Sampled adversarial examples via average approximation over the substitute model with multiple inferences.

Main Results:

• The proposed method achieves state-of-the-art performance in no-box adversarial attacks.
• Demonstrated successful adversarial attacks using solely diffusion model-generated data.
• Validated effectiveness on the challenging ImageNet dataset.

Conclusions:

• Diffusion models can be effectively utilized for sophisticated adversarial attacks.
• The developed no-box attack method offers a practical solution for real-world threat scenarios.
• This research highlights the dual nature of diffusion models as powerful generative tools and potential security risks.