jonasricker / aeroblade

[CVPR2024] AEROBLADE: Training-Free Detection of Latent Diffusion Images Using Autoencoder Reconstruction Error
https://openaccess.thecvf.com/content/CVPR2024/html/Ricker_AEROBLADE_Training-Free_Detection_of_Latent_Diffusion_Images_Using_Autoencoder_Reconstruction_CVPR_2024_paper.html
40 stars 6 forks source link
autoencoder cvpr deepfake-detection diffusion-models latent-diffusion midjourney stable-diffusion

[CVPR2024] AEROBLADE: Training-Free Detection of Latent Diffusion Images Using Autoencoder Reconstruction Error

Jonas Ricker, Denis Lukovnikov, Asja Fischer
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

header

Setup

Create a virtual environment and run

pip install -r requirements.txt
pip install -e .

(tested with Python 3.10).

Quickstart

If you simply want to use AEROBLADE for detection, run

python scripts/run_aeroblade.py --files-or-dirs path/to/img.png path/to/dir

Calling the script without any arguments will use the images in example_images. By default, it computes the reconstructions using the AEs from Stable Diffusion 1, Stable Diffusion 2, and Kandinsky 2.1 and measures the distance using the second LPIPS layer, which performed best in our experiments. Use -h to learn more about the available options.

The computed distances are printed and saved to aeroblade_output/distances.csv. Note that we save the negative distances, which is why the best AE is denoted by max.

Reproducing our Experiments

Here we provide the commands to reproduce our experimental results. Note the following:

Data

Our data can be downloaded from Zenodo. Extract the .zip file and place the data directory inside the root directory of this repository.

Generated Images All generated images are readily available in data/raw/generated. For Stable Diffusion and Kandinsky, the directory name also contains the CLIP model used to extract the prompts, e.g., CompVis-stable-diffusion-v1-1-ViT-L-14-openai. We also provide the extracted prompts in data/raw/prompts.

Real Images Since the real images from LAION-5B are subject to copyright, we cannot distribute them (see FAQ). Therefore, you will have to reconstruct the real dataset from the image URLs provided in data/real/real_metadata.parquet. The following command will attempt to download the images from their respective URLs to tmp/laion:

img2dataset --url_list data/raw/real/real_metadata.parquet --input_format "parquet" --url_col "URL" --caption_col "TEXT" --output_folder tmp/laion --resize_mode "center_crop" --min_image_size 512 --max_image_area 589824 --image_size 512 --encode_format "png" --encode_quality 6

To ensure that the filenames correspond to the extracted prompts, you can run

python scripts/rename_real_images.py

The renamed files will be saved to data/raw/real.

Note: Images from LAION tend to become unavailable over time, which is why your dataset will likely be incomplete. As a remedy, our scripts have the option to use pre-computed results for real images, which is detailed below. You can also run the evaluations with a reduced number of images or use different real images.

5.2. Evaluation of Detection Performance

python experiments/01_detect.py

To use pre-computed distances of real images add --precomputed-real-dist data/precomputed/01_default_real_dist.pickle.

5.4. Qualitative Image Analysis with AEROBLADE

Relation Between Image Complexity and Reconstruction Error

python experiments/02_analyze_patches.py

To use pre-computed distances and complexities of real images add --precomputed-real-dist data/precomputed/02_default_real_dist.pickle --precomputed-real-compl data/precomputed/02_default_real_compl.pickle.

5.5. Additional Analyses

Robustness to Perturbations

python experiments/01_detect.py --experiment-id robustness --amount 250 --transforms clean jpeg_90 jpeg_80 jpeg_70 jpeg_60 jpeg_50 blur_1.0 blur_2.0 blur_3.0 blur_4.0 blur_5.0 crop_0.9 crop_0.8 crop_0.7 crop_0.6 crop_0.5 noise_0.05 noise_0.1 noise_0.15 noise_0.2 noise_0.25

To use pre-computed distances of real images add --precomputed-real-dist data/precomputed/01_robustness_real_dist.pickle.

Exploring Different Distance Metrics

python experiments/01_detect.py --experiment-id distance_metric_ablation --distance-metrics lpips_vgg_0 lpips_alex_0 lpips_squeeze_0 psnr ssimc ms_ssim dists

To use pre-computed distances of real images add --precomputed-real-dist data/precomputed/01_distance_metric_ablation_real_dist.pickle.

Using Deeper Reconstructions

python experiments/03_deeper_reconstructions.py --experiment-id deeper_sd15 --real-dir data/raw/real --fake-dir data/raw/generated/runwayml-stable-diffusion-v1-5-ViT-L-14-openai --repo-id runwayml/stable-diffusion-v1-5
python experiments/03_deeper_reconstructions.py --experiment-id deeper_sd21 --real-dir data/raw/real --fake-dir data/raw/generated/stabilityai-stable-diffusion-2-1-base-ViT-H-14-laion2b_s32b_b79k --repo-id stabilityai/stable-diffusion-2-1-base

To use pre-computed distances of real images add --precomputed-real-dist data/precomputed/03_deeper_sd15_real_dist.pickle and --precomputed-real-dist data/precomputed/03_deeper_sd21_real_dist.pickle, respectively.