Towards Quantitative Evaluation of Explainable AI Methods for Deepfake Detection

PyTorch Implementation [Paper] [DOI] [Cite]

• From "Towards Quantitative Evaluation of Explainable AI Methods for Deepfake Detection", Proc. ACM Int. Workshop on Multimedia AI against Disinformation (MAD’24) at the ACM Int. Conf. on Multimedia Retrieval (ICMR’24), Thailand, June 2024.
• Written by Konstantinos Tsigos, Evlampios Apostolidis, Spyridon Baxevanakis, Symeon Papadopoulos and Vasileios Mezaris.
• This software can be used to evaluate the performance of five explanation approaches from the literature (GradCAM++, RISE, SHAP, LIME, SOBOL), on explaining the output of a state-of-the-art model (based on Efficient-Net) for deepfake detection. Our evaluation framework assesses the ability of an explanation method to spot the regions of a fake image with the biggest influence on the decision of the deepfake detector, by examining the extent to which these regions can be modified through a set of adversarial attacks, in order to flip the detector's prediction or reduce its initial prediction.

Dependencies

The code was developed, checked and verified on an Ubuntu 20.04.6 PC with an NVIDIA RTX 4090 GPU and an i5-12600K CPU. All dependencies can be found inside the environment.yml file, which can be used to set up the necessary conda enviroment.

Data

The data for re-producing our experiments are the videos from the test split of the FaceForensics++ dataset. This dataset contains 1000 original videos and 4000 fake videos created using one of the following four classes of AI-based manipulation (1000 per class): "FaceSwap" (FS), "DeepFakes" (DF), "Face2Face" (F2F), and "NeuralTextures" (NT). The videos of the FS class were created via a graphics-based approach that transfers the face region from a source to a target video. The videos of the DF class were produced using autoencoders to replace a face in a target sequence with a face in a source video or image collection. The videos of the F2F class were obtained by a facial reenactment system that transfers the expressions of a source to a target video while maintaining the identity of the target person. The videos of the NT class were generated by modifying the facial expressions corresponding to the mouth region, using a patch-based GAN-loss. The dataset is divided into training, validation, and test sets, comprised of 720, 140 and 140 videos, respectively.

To re-create the database file that we used in our experiments, please follow the steps below:

1. Download the entire FaceForensics++ dataset
2. Run the following script to preprocess the raw data:

python data/preprocess_ff.py prepro -r RAW_DATA_PATH -tr PREPROCESSED_DATA_PATH -d cuda:0 -mdcsv RAW_DATA_PATH/dataset_info.csv -orig

where RAW_DATA_PATH is the path to the downloaded FF++ dataset and PREPROCESSED_DATA_PATH is the path to save the preprocessed data. The script will create a new file faceforensics_frames.csv containing the paths to the preprocessed frames.

3. Create a new LMDB database by running the following script:

python data/lmdb_storage.py add-csv -csv faceforensics_frames.csv -h -pc relative_path -d ./data/xai_test_data.lmdb -ms 21474836480 -v -b PREPROCESSED_DATA_PATH

where faceforensics_frames.csv is the file created in the previous step and PREPROCESSED_DATA_PATH is the path to the preprocessed data. The script will create a new LMDB database xai_test_data.lmdb containing the preprocessed frames. The -ms flag specifies the maximum size of the database in bytes, default is 20GB.

Trained model

The employed model (called ff_attribution) was trained for multiclass classification on the FaceForensics++ dataset. It outputs a probability for each of the 5 classes (0, 1, 2, 3, 4), corresponding to "Real", "NeuralTextures", "Face2Face", "DeepFakes" and "FaceSwap", respectively.

Model characteristics

Model	ff_attribution
Task	multiclass
Arch.	efficientnetv2_b0
Type	CNN
No. Params	7.1M
No. Datasets	1
Input	(B, 3, 224, 224)
Output	(B, 5)

Performance (FF++ test set)

Metric	Value
MulticlassAccuracy	0.9626
MulticlassAUROC	0.9970
MulticlassF1Score	0.9627
MulticlassAveragePrecision	0.9881

Evaluation and visualization

To evaluate the explanation method(s) using our framework, run the evaluate.py script.

To evaluate the explanation method(s) using both our framework and the one from Gowrisankar et al. (2024) (which was used as the basis for developing our framework), and compare the results (as we did in Tables 3 and 4 of our paper), run the evaluate_pipelines_comparison.py script.

To visualize the created explanation by an explanation method, for a specific image of the dataset, run the visualize.py script.

Running parameters and evaluation results

The default parameters and the available options for running the above scripts, are listed below:

Parameter name	File	Description	Default Value	Options
evaluation_explanation_methods	evaluate.py evaluate_pipelines_comparison.py	Explanation method(s) to evaluate	'All'	'All', 'GradCAM++', 'RISE', 'SHAP', 'LIME', 'SOBOL'
explanation_method	visualize.py	Explanation method to explain the image.	'LIME'	'GradCAM++', 'RISE', 'SHAP', 'LIME', 'SOBOL'
dataset_example_index	visualize.py	Index of the image in the database	'random'	'random', integer between [0,13837]

The evaluation results are printed onto the console and saved into a CSV file which is stored within the results folder, created at the explanation path. To eliminate the need to run the evaluation process from the beginning in case of an unexpected error, the evaluation results for each test image are accumulated into an NPY file, that is used as a checkpoint and is also located at the results folder.

Citation

If you find our work, code or pretrained models, useful in your work, please cite the following publication:

K. Tsigos, E. Apostolidis, S. Baxevanakis, S. Papadopoulos, V. Mezaris, "Towards Quantitative Evaluation of Explainable AI Methods for Deepfake Detection", Proc. ACM Int. Workshop on Multimedia AI against Disinformation (MAD’24) at the ACM Int. Conf. on Multimedia Retrieval (ICMR’24), Thailand, June 2024.

The accepted version of this paper is available on ArXiv at: https://arxiv.org/abs/2404.18649

BibTeX:

@INPROCEEDINGS{Tsigos2024,
    author    = {Tsigos, Konstantinos, and Apostolidis, Evlampios and Baxevanakis, Spyridon and Papadopoulos, Symeon and Mezaris, Vasileios},
    title     = {Towards Quantitative Evaluation of Explainable AI Methods for Deepfake Detection},
    year      = {2024},
    publisher = {Association for Computing Machinery},
    address   = {New York, NY, USA},
    booktitle = {Proceedings of the 3rd ACM International Workshop on Multimedia AI against Disinformation},
    location  = {Phuket, Thailand},
    series    = {MAD '24}
}

License

This code is provided for academic, non-commercial use only. Please also check for any restrictions applied in the code parts and datasets used here from other sources. For the materials not covered by any such restrictions, redistribution and use in source and binary forms, with or without modification, are permitted for academic non-commercial use provided that the following conditions are met:

Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation provided with the distribution.

This software is provided by the authors "as is" and any express or implied warranties, including, but not limited to, the implied warranties of merchantability and fitness for a particular purpose are disclaimed. In no event shall the authors be liable for any direct, indirect, incidental, special, exemplary, or consequential damages (including, but not limited to, procurement of substitute goods or services; loss of use, data, or profits; or business interruption) however caused and on any theory of liability, whether in contract, strict liability, or tort (including negligence or otherwise) arising in any way out of the use of this software, even if advised of the possibility of such damage.

Acknowledgement

This work was supported by the EU Horizon Europe and Horizon 2020 programmes under grant agreements 101070190 AI4TRUST and 951911 AI4Media, respectively.