New methods improve diffusion model privacy audits by separating memorization from image difficulty

Researchers have developed new methods to improve the accuracy of membership inference attacks (MIAs) on diffusion models, which are used to detect if specific images were part of a model's training data. The standard approach of using reconstruction loss is flawed because image difficulty can be mistaken for memorization. The new techniques, Reference-Model Calibration (RMC) and Temporal Self-Calibration (TSC), address this by calibrating for image difficulty across training and diffusion times, respectively. RMC compares a fine-tuned model's loss to an earlier checkpoint, while TSC uses a single model to compare loss across diffusion timesteps, leveraging the observation that image difficulty is consistent over time while memorization is localized. These methods significantly outperform existing state-of-the-art attacks and generalize across different diffusion model parameterizations. AI

IMPACT Enhances privacy auditing capabilities for diffusion models, potentially impacting how training data is managed and secured.

RANK_REASON Academic paper detailing novel methods for membership inference attacks on diffusion models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Hugging Face Daily Papers →

• Diffusion Models
• Hugging Face
• Membership Inference Attacks
• MichaelJoseph777/TSC
• Reference-Model Calibration
• Temporal Self-Calibration

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →