LoFIT
A membership inference method for diffusion models that distinguishes training from non-training samples by quantifying the initial response of the denoising loss along the normalized conditional prediction shift direction. It requires no model gradients and relies only on forward queries.
Timeline (last 90 days)
LoFIT (Local Conditional Fit) quantifies the initial response of the denoising loss along the normalized prediction shift and aggregates it over multiple text conditions and diffusion states, requiring no gradients from the target model and only forward queries to the denoiser.
Evaluated on four datasets, LoFIT achieves better membership discrimination on both fine-tuned models and pre-trained models.
In the early stage of fine-tuning, LoFIT can already distinguish members, while the naive denoising loss baseline remains near random level.
LoFIT outperforms existing methods in the black-box setting and does not require access to the model's internal gradients.
LoFIT has so far been validated only on four datasets and mainly targets text-to-image diffusion models; future work could explore more modalities (such as video diffusion) and larger-scale models.
All Facts (5)
LoFIT (Local Conditional Fit) quantifies the initial response of the denoising loss along the normalized prediction shift and aggregates it over multiple text conditions and diffusion states, requiring no gradients from the target model and only forward queries to the denoiser.
50%UnverifiedEvaluated on four datasets, LoFIT achieves better membership discrimination on both fine-tuned models and pre-trained models.
50%UnverifiedIn the early stage of fine-tuning, LoFIT can already distinguish members, while the naive denoising loss baseline remains near random level.
50%UnverifiedLoFIT outperforms existing methods in the black-box setting and does not require access to the model's internal gradients.
50%UnverifiedLoFIT has so far been validated only on four datasets and mainly targets text-to-image diffusion models; future work could explore more modalities (such as video diffusion) and larger-scale models.
50%