How to Trace Latent Generative Model Generated Images without Artificial Watermark?
Authors: Zhenting Wang, Vikash Sehwag, Chen Chen, Lingjuan Lyu, Dimitris N. Metaxas, Shiqing Ma
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments on the state-of-the-art latent generative models, such as Stable Diffusion, show that our method can distinguish the images generated by the inspected model and other images with a high accuracy and efficiency. Our findings suggest the intriguing possibility that today s latent generative generated images are naturally watermarked by the decoder used in the source models. |
| Researcher Affiliation | Collaboration | 1Rutgers University 2Sony AI 3University of Massachusetts at Amherst. |
| Pseudocode | Yes | Algorithm 1 Origin Attribution by Gradient-based Inversion with Encoder-based Initialization |
| Open Source Code | Yes | Code: https://github. com/Zhenting Wang/Latent Tracer. [...] Our code can be found at https: //github.com/Zhenting Wang/Latent Tracer. |
| Open Datasets | Yes | We first randomly sample 500 images and their corresponding text captions in the LAION dataset (Schuhmann et al., 2022), which is the training dataset for many state-of-the-art text-to-image latent generative models such as Stable Diffusion (Rombach et al., 2022). [...] For Quantized Auto-encoder, we use a VQ-VAE (Van Den Oord et al., 2017) model trained on the CIFAR-10 dataset (Krizhevsky et al., 2009). [...] For Autoregressive Model, we use a Vi T-VQGAN (Yu et al., 2021a) model trained on the Image Net dataset (Russakovsky et al., 2015). |
| Dataset Splits | No | The paper does not explicitly provide training, validation, or test dataset splits for the proposed LATENTTRACER method, as the method itself is not a trainable model in the conventional sense. It describes how images were generated or selected for evaluation (e.g., '500 images for each model', 'randomly sampled images from LAION'), but these are evaluation sets, not formal dataset splits for model training. |
| Hardware Specification | Yes | We conducted all experiments on a Ubuntu 20.04 server equipped with 8 A100 GPUs (one experiment/GPU). |
| Software Dependencies | Yes | Our method is implemented with Python 3.10 and Py Torch 2.0. |
| Experiment Setup | Yes | In our implementation, we set 100 as the maximum step. [...] In detail, it utilizes the model M to generate N (defaulting to 100 in this paper) images from randomly selected prompts. [...] Following Wang et al. (2023d), we use Grubbs Hypothesis Testing (Grubbs, 1950) to determine the threshold: [...] where α is the significance level of the hypothesis testing (i.e., 0.05 by default in this paper). |