Latent Code Augmentation Based on Stable Diffusion for Data-Free Substitute Attacks.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Cybersecurity

Background:

• Black-box substitute attacks lack target model training data.
• Current methods use generative adversarial networks (GANs), facing low efficiency and generation quality issues.
• GANs require retraining for each target model, limiting scalability.

Purpose of the Study:

• To propose a novel data-free substitute attack scheme using diffusion models (DMs) for improved efficiency and accuracy.
• To address the domain distribution and sample variation challenges of DM-generated data.
• To enhance the training of substitute models that closely mimic target models.

Main Methods:

• Utilized stable diffusion (SD), a type of diffusion model (DM), for data generation.
• Introduced latent code augmentation (LCA) to align SD-generated data with target model distributions.
• Augmented latent codes of inferred member data to guide SD, ensuring data diversity and discriminative criteria adherence.

Main Results:

• Latent code augmentation (LCA) facilitated stable diffusion (SD) to generate high-quality, diverse data aligned with target model distributions.
• Substitute models trained with LCA-guided data achieved higher attack success rates (ASRs).
• The proposed method required fewer query budgets compared to existing GANs-based schemes.

Conclusions:

• The novel data-free substitute attack scheme based on stable diffusion (SD) and latent code augmentation (LCA) significantly improves efficiency and accuracy.
• LCA effectively overcomes the limitations of GANs in data-free substitute attacks.
• This approach offers a more efficient and effective method for training substitute models in black-box scenarios.