Deepfake detection
4 likes27,862 views
Deepfake detection models require clean training data to generalize well. The document discusses preprocessing training data by filtering out false detections from face extraction. This improved log loss error on evaluation datasets for models trained with the preprocessed data. However, deepfake detection remains challenging due to limited generalization, overfitting, and the broad scope of possible manipulations. The importance of preprocessing training data and methods to address challenges are discussed.
![More details at TTO 2020
Charitidis, P., Kordopatis-Zilos, G., Papadopoulos, S., & Kompatsiaris, Y.
(2020). Investigating the impact of preprocessing and prediction
aggregation on the DeepFake detection task. Proceedings of the
Conference for Truth and Trust Online (TTO) [to appear],
https://arxiv.org/abs/2006.07084
https://truthandtrustonline.com/](https://image.slidesharecdn.com/deepfake-detection-weverify-oct2020-201020143809-201020201745/85/Deepfake-detection-18-320.jpg)
Deepfake detection
- 1. DeepFake Detection: The Importance of Training Data Preprocessing and Practical Considerations Dr. Symeon (Akis) Papadopoulos – @sympap MeVer Team @ Information Technologies Institute (ITI) / Centre for Research & Technology Hellas (CERTH) Joint work with Polychronis Charitidis, George Kordopatis-Zilos and Yiannis Kompatsiaris AI4Media Workshop on GANs for Media Content Generation, Oct 1, 2020 Media Verification (MeVer)
- 2. DeepFakes • Content, generated by AI, that seems authentic to human eye • Most common form: generation and manipulation of human face Source: https://en.wikipedia.org/wiki/Deepfake Source: https://www.youtube.com/watch?v=iHv6Q9ychnA Source: Media Forensics and DeepFakes: an overview
- 3. Manipulation types Facial manipulations can be categorised in four main different groups: • Entire face synthesis • Attribute manipulation • Identity swap • Expression swap Source: DeepFakes and Beyond: A Survey of Face Manipulation and Fake Detection (Tolosana et al., 2020) Tolosana, R., et al. (2020). Deepfakes and beyond: A survey of face manipulation and fake detection. arXiv preprint arXiv:2001.00179. Verdoliva, L. (2020). Media forensics and deepfakes: an overview. arXiv preprint arXiv:2001.06564. Mirsky, Y., & Lee, W. (2020). The Creation and Detection of Deepfakes: A Survey. arXiv preprint arXiv:2004.11138.
- 4. WeVerify Project • WeVerify aims at detecting disinformation in social media and expose misleading and fabricated content • Partners: Univ. Sheffield, OntoText, ATC, DW, AFP, EU DisinfoLab, CERTH • A key outcome is a platform for collaborative content verification, tracking, and debunking • Currently, we are developing a deepfake detection service for the WeVerify platform • Participation in DeepFake Detection Challenge https://weverify.eu/
- 5. DeepFake Detection Challenge • Goal: detect videos with facial or voice manipulations • 2,114 teams participated in the challenge • Log Loss error evaluation on public and private validation sets • Public evaluation contained videos with similar transformations as the training set • Private evaluation contained organic videos and videos with unknown transformations from the Internet • Our final standings: • public leaderboard: 49 (top 3%) with 0.295 Log Loss error • private leaderboard: 115 (top 5%) with 0.515 Log Loss error Source: https://www.kaggle.com/c/deepfake-detection-challenge
- 6. DeepFake Detection Challenge - dataset • Dataset of more than 110k videos • Approx. 20k REAL and the rest are FAKE • FAKE videos generated from the REAL • Models used: • DeepFake AutoEncoder (DFAE) • Morphable Mask faceswap (MM/NN) • Neural Talking Heads (NTH) • FSGAN • StyleGAN Dolhansky, B., Bitton, J., Pflaum, B., Lu, J., Howes, R., Wang, M., & Ferrer, C. C. (2020). The DeepFake Detection Challenge Dataset. arXiv preprint arXiv:2006.07397.
- 7. Dataset preprocessing - Issues • Face dataset quality depends on face extraction accuracy (Dlib, mtcnn, facenet-pytorch, Blazeface) • Generally all face extraction libraries generate a number of false positive detections • Manual tuning can improve the quality of the generated dataset Deep learning model Face extraction Frame extraction Video corpus
- 8. Noisy data creeping in the training set • Extracting faces with 1 fps from Kaggle DeepFake Detection Challenge dataset videos using pytorch implementation of MTCNN face detection • Observation: False detections are less compared to true detections in a video
- 9. Our “noise” filtering approach • Compute face embeddings for each detected face in video • Similarity calculation between all face embeddings in a video → similarity graph construction • Nodes represent faces and two faces are connected if their similarities are greater than 0.8 (solid lines) • Drop components smaller than N/2 (e.g. component 2) • N is the number of frames that contain face detections (true or false).
- 10. Advantages • Simple and fast procedure • No need for manual tuning of the face extraction settings • Clusters of distinct faces in cases of multiple persons in the video • This information can be utilized in various ways (e.g. predictions per face) Faces extracted from multiple video frames Component 1 Component 2
- 11. Experiments • We trained multiple DeepFake detection models on the DFDC dataset with and without (baseline) our proposed approach • Three datasets: a) Celeb-DF, b) FaceForensics++, c) DFDC subset • For evaluation we examined two aggregation approaches • avg: prediction is the average of all face predictions • face: prediction is the max prediction among different avg face predictions • Results for the EfficientNet-B4 model in terms of Log loss error: Pre- processing CelebDF FaceForensics++ DFDC avg face avg face avg face baseline 0,510 0,511 0,563 0,563 0,213 0,198 proposed 0,458 0,456 0,497 0,496 0,195 0,173
- 12. Our DFDC Approach - details • Applied proposed preprocessing approach to clean the generated face dataset • Face augmentation: • Horizontal & vertical flip, random crop, rotation, image compression, Gaussian & motion blurring, brightness, saturation & contrast transformation • Trained three different models: a) EfficientNet-B3, b) EfficientNet-B4, c) I3D* • Models trained on face level: • I3d trained with 10 consecutive face images exploiting temporal information. • EfficientNet models trained on single face images • Per model: • Added two dense layers with dropout after the backbone architecture with 256 and 1 units • Used the sigmoid activation for the last layer * ignoring the optical flow stream
- 13. Our DFDC approach – inference pre-processing model inference post-processing
- 14. Lessons from other DFDC teams • Most approaches ensemble multiple EfficientNet architectures (B3-B7) and some of them were trained on different seeds • ResNeXT was another architecture used by a top-performing solutions combined with 3D architectures such as I3D, 3D ResNet34, MC3 & R2+1D • Several approaches increased the margin of the detected facial bounding box to further improve results. • We used an additional margin of 20% but other works proposed a higher proportion. • To improve generalization: • Domain-specific augmentations: a) half face removal horizontally or vertically, b) landmark (eyes, nose, or mouth) removal • Mixup augmentations
- 15. Practical challenges • Limited generalization • This observation applies to most submissions. The winning team scored 0.20336 in public validation and only 0.42798 in the private (Log Loss) • Overfitting • The best submission in the public leaderboard scored 0.19207 but in the private evaluation the error was 0.57468, leading to the 904-th position! • Broad problem scope • The term DeepFake may refer to every possible manipulation and generation • Constantly increasing manipulation and generation techniques • A detector is only trained with a subset of these manipulations
- 16. DeepFake Detection in the Wild • Videos in the wild usually contain multiple scenes • Only a subset of these scenes may contain DeepFakes • Detection process might be slow for multi-shot videos (even short ones) • Low quality videos • Low quality faces tend to fool classifiers • Small detected and fast-moving faces • Usually lead to noisy predictions • Changes in the environment • Moving obstacles in front of the faces • Changes in lighting
- 17. DeepFake Detection Service @ WeVerify https://www.youtube.com/watch?v=cVljNV V5VPw&ab_channel=TheFakening
- 18. More details at TTO 2020 Charitidis, P., Kordopatis-Zilos, G., Papadopoulos, S., & Kompatsiaris, Y. (2020). Investigating the impact of preprocessing and prediction aggregation on the DeepFake detection task. Proceedings of the Conference for Truth and Trust Online (TTO) [to appear], https://arxiv.org/abs/2006.07084 https://truthandtrustonline.com/
- 19. Thank you! Dr. Symeon Papadopoulos papadop@iti.gr @sympap Media Verification (MeVer) https://mever.iti.gr/ @meverteam https://ai4media.eu/ https://weverify.eu/