Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
MYOPIA: Protecting Face Privacy from Malicious Personalized Text-to-Image Synthesis via Unlearnable Examples
Authors: Zhihao Wu, Yushi Cheng, Tianyang Sun, Xiaoyu Ji, Wenyuan Xu
AAAI 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Evaluation on two face datasets, i.e., VGGFace2 and Celeb A-HQ, with various model versions illustrates the effectiveness of our approach in preserving personal privacy. Furthermore, our method showcases robust transferability across diverse model versions and demonstrates resilience against various image pre-processing techniques. Extensive experiments demonstrate that our methods effectively prevent personalized T2I models from capturing genuine facial features of the target individual, exhibiting strong portability and robustness. |
| Researcher Affiliation | Academia | 1USSLAB, Zhejiang University 2ZJU-UIUC Institute, Zhejiang University EMAIL, EMAIL |
| Pseudocode | No | The paper mentions 'A detailed implementation of the algorithm can be found in the Appendix.' but does not include any pseudocode or algorithm blocks in the main text. |
| Open Source Code | Yes | Code https://github.com/Zhihao Wu95/myopia |
| Open Datasets | Yes | We evaluate MYOPIA on two well-known face datasets VGGFace2 (Cao et al. 2018) and Celeb A-HQ (Liu et al. 2015). |
| Dataset Splits | No | The paper mentions: 'Each identity comprises two subsets: a target-protected image set and a clean image set for reference. Both subsets consist of four images resized to 512 512.' and discusses 'poisoning rates' (e.g., '4/4', '3/4') which describe the ratio of protected images to all fine-tuning images. It does not provide specific train/test/validation dataset splits for the overall datasets, but rather how images are used per identity during fine-tuning. |
| Hardware Specification | Yes | For both the clean image set and the protected image set, we train a Stable Diffusion model using the Dream Booth fine-tuning technique on an NVIDIA H800 GPU (80GB). |
| Software Dependencies | No | The paper mentions using 'Stable Diffusion model version 2.1' and 'Dream Booth fine-tuning technique'. While it specifies the version of the model used, it does not list general software dependencies like programming languages or libraries (e.g., Python, PyTorch) with their specific version numbers required for reproducibility. |
| Experiment Setup | Yes | The training parameters includes a learning rate of 5e-7, a batch size of 2, and 1000 training steps. We choose the Sable Diffusion model version 2.1 as the default model. During training, we follow the same setting as Dream Booth, utilizing the instance prompt a photo of sks person and the prior prompt a photo of person . We generate the unlearnable perturbation by alternating optimization with default parameters: the iteration K set to 10, the PGD steps N of 100, the step size α of 0.005, the JND weight λ of 0.985, and the perturbation bounded at L -norm 0.5. |