Abstract
Membership inference attacks (MIAs) on diffusion models have emerged as potential evidence of unauthorized data usage in training pre-trained diffusion models. These attacks aim to detect the presence of specific images in training datasets of diffusion models. Our study delves into the evaluation of state-of-the-art MIAs on diffusion models and reveals critical flaws and overly optimistic performance estimates in existing MIA evaluation. We introduce CopyMark, a more realistic MIA benchmark that distinguishes itself through the support for pre-trained diffusion models, unbiased datasets, and fair evaluation pipelines. Through extensive experiments, we demonstrate that the effectiveness of current MIA methods significantly degrades under these more practical conditions. Based on our results, we alert that MIA, in its current state, is not a reliable approach for identifying unauthorized data usage in pre-trained diffusion models. To the best of our knowledge, we are the first to discover the performance overestimation of MIAs on diffusion models and present a unified benchmark for more realistic evaluation.
Introduction
As generative AI faces increasing legal scrutiny regarding copyright, Membership Inference Attacks (MIAs) have been proposed as a "smoking gun" to prove unauthorized data usage. In theory, MIAs analyze a model's behavior to determine if a specific data point was part of its training set. However, current MIA evaluations often suffer from two major defects: over-training on toy models and dataset shift between member and non-member sets. These factors lead to inflated performance metrics that do not translate to real-world, professional models.
The CopyMark Benchmark
To address these limitations, we developed CopyMark, the first unified benchmark designed for realistic MIA evaluation. CopyMark addresses previous flaws through:
- A Two-Stage Pipeline: Separating attack tuning (validation) from performance measurement (blind testing) to ensure fair evaluation.
- Support for Pre-trained Models: Integrating with industry-standard models like Stable Diffusion v1.5 and SDXL, which are typically trained for only one or two epochs.
- Unbiased Datasets: Utilizing setups where members and non-members are drawn from the same high-quality distribution, such as LAION-2B and HakuBooru.
Analysis of the "Dataset Classifier" Trap
A key finding of our study is that many MIAs act as "dataset classifiers." When non-member images come from a different source than the training data, the attack learns to distinguish the two sources based on style or semantics rather than actual membership.
When this distribution shift is removed, the performance of these attacks drops precipitously, revealing their lack of true membership detection capability.
Key Results and Discussion
Our extensive experiments on CopyMark show that current MIAs are largely ineffective on professional models.
- Loss-based Failure: Methods like SecMI and PIA perform no better than random guessing on unshifted, realistic datasets.
- Classifier-based Overfitting: While methods like GSA appear strong on validation sets, they fail to generalize to blind test sets, with False Positive Rates often exploding from 1% to over 40%.
- The 1-Epoch Shield: Professional models trained for only one epoch do not exhibit the memorization patterns that MIAs rely on.
Conclusion
MIAs, in their current state, are not reliable tools for AI copyright lawsuits. They provide binary indicators that are easily fooled by dataset shifts and fail to meet the legal standard of "substantial similarity." Future work must focus on identifying features truly unique to training data that don't just reflect general dataset style.
Summary updated