Slides accompanying an online webinar on DeepFake Detection and a hands-on demonstration of the MeVer DeepFake Detection service. The webinar is supported by the US-Paris Tech Challenge award for our work on the InVID-WeVerify plugin.
Symeon PapadopoulosResearcher at CERTH-ITI, Co-founder at infalia um infalia
DeepFake Detection: Challenges, Progress and Hands-on Demonstration of Technology
1. DeepFake Detection
Challenges, Progress and
Hands-on Demonstration of Technology
Dr. Symeon (Akis) Papadopoulos – @sympap
Dr. Nikos Sarris - @nikossarris
MeVer Team @ Information Technologies Institute (ITI) /
Centre for Research & Technology Hellas (CERTH)
Online Webinar, Dec 16th 2021
Media Verification
(MeVer)
2. DeepFakes
Content, generated by deep neural
networks, that seems authentic to
human eye
Four main types of face DeepFakes:
a) Entire face synthesis, b) Attribute
manipulation, c) Identity swap,
d) 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.
Information Fusion, 64, 131-148.
Verdoliva, L. (2020). Media forensics and deepfakes:
an overview. IEEE Journal of Selected Topics in
Signal Processing, 14(5), 910-932.
Mirsky, Y., & Lee, W. (2021). The creation and
detection of deepfakes: A survey. ACM Computing
Surveys (CSUR), 54(1), 1-41.
reenactment
replacement
editing
generation
3. Gaining popularity
Nguyen, T. T., et al. (2019). Deep learning for
deepfakes creation and detection. arXiv preprint
arXiv:1909.11573, 1.
Ajder, H., et al. (2019).The State of DeepFakes:
Landscape, Threats and Impact. Report by
DeepTraceLabs/Sensity.
4. Potential Risks and Harms
Tackling deepfakes in European policy, Panel for the Future of Science and Technology,
Scientific Foresight Unit (STOA), July 2021
5. DeepFakes and Politics
One week after the video’s release, Gabon’s military
attempted an ultimately unsuccessful coup—the country’s
first since 1964—citing the video’s oddness as proof
something was amiss with the president.
https://www.motherjones.com/politics/2019/03/deepfake
-gabon-ali-bongo/
Mr Nguyen said he could not rule out the video being a
‘deepfake’, a term for the fairly new artificial intelligence
based technology which involves machine learning
techniques to superimpose a face on a video.
https://www.sbs.com.au/news/a-gay-sex-tape-is-threatening-
to-end-the-political-careers-of-two-men-in-malaysia
6. DeepFake Quality Rapidly Improving
https://twitter.com/goodfellow_ian/status/1084973596236144640
2021
Masood, M., Nawaz, M., Malik, K. M., Javed, A., & Irtaza, A. (2021). Deepfakes Generation and Detection: State-of-
the-art, open challenges, countermeasures, and way forward. arXiv preprint arXiv:2103.00484.
Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., & Aila, T. (2021, May). Alias-free generative
adversarial networks. In Thirty-Fifth Conference on Neural Information Processing Systems.
7. A New Level of Realism
• Created by Chris Ume, a VFX specialist
• Not detected by any of the commercial
deepfake detection services
• Not discernible by human inspection
• Potential for misleading
but to date barriers are still high
• a lot of expertise, skill and time
• an impersonator who looks like the target
(Miles Fisher)
https://www.theverge.com/2021/3/5/22314980/tom-cruise-
deepfake-tiktok-videos-ai-impersonator-chris-ume-miles-fisher
8. Common DF Neural Network Architectures
Mirsky, Y., & Lee, W. (2021). The Creation and Detection of Deepfakes: A Survey. ACM Computing Surveys, 54(1), 1-41.
9. DeepFake Creation Pipeline and Tools
Mirsky, Y., & Lee, W. (2021). The Creation
and Detection of Deepfakes: A Survey. ACM
Computing Surveys, 54(1), 1-41.
faceswap.dev
https://github.com/iperov/DeepFaceLab
zaodownload.com malavida.com/en/soft/fakeapp
hey.reface.ai
facemagic.ai
https://generated.photos/face-generator
10. Signs of a DeepFake (in 2021)
• Different kinds of
artifacts
• Blurry areas around lips,
hair, earlobs
• Lack of symmetry
• Lighting inconsistencies
• Fuzzy background
• Flickering (in video)
https://apnews.com/article/bc2f19097a4c4fffaa00de6770b8a60d
11. DF Landscape: Detection Approaches
PHYSIOLOGICAL
SIGNALS
Blinking
information
Corneal specular
highlights
Photo-
plethysmography
ARTIFACT BASED
DETECTION
3D head pose
features
Limited resolution /
blurring
Local artifacts
(eyes, teeth, etc.)
Face X-Ray
(blending artifacts)
DEEP LEARNING
ARCHITECTURES
MesoNet XceptionNet
Capsule
Networks
Recurrent
Convolutions
FREQUENCY
DOMAIN
Local frequency
statistics
Spectral
distribution
Two-stream
approaches
Attention Nets
(Transformers)
12. The DF Battleground
• DeepFake generation and detection
offer a naturally adversarial setting.
• A recent survey (Feb 2021) analyzed
70 generation and 108 detection
methods and linked them if a
detection method tried to detect
media from a given generator.
• Analysis indicates the fast evolution
of this field.
Juefei-Xu, F., Wang, R., Huang, Y., Guo, Q., Ma, L., & Liu, Y.
(2021). Countering Malicious DeepFakes: Survey,
Battleground, and Horizon. arXiv preprint arXiv:2103.00218.
14. 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
Source: https://www.kaggle.com/c/deepfake-detection-challenge
15. Performance of SotA methods on DFDC
The DFDC highlights the
generalization challenge
faced by SotA methods.
public set
hidden set
Kim, M., Tariq, S., & Woo, S. S. (2021). FReTAL: Generalizing Deepfake Detection using Knowledge Distillation and Representation Learning. In
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 1001-1012).
Accuracy in cross forgery experiments (FF++ HQ)
Method DF F2F FS NT
Xception (DF) 99.41 56.05 49.93 66.32
Xception (F2F) 68.55 98.64 50.55 54.81
Xception (FS) 49.89 54.15 98.36 50.74
Xception (NT) 50.05 57.49 50.01 99.88
Accuracy in cross dataset experiments
Method FF++ HQ CELEB-DF
Xception (FF++ HQ) 95.60 73.01
16. The MeVer DeepFake Detection Service
• R&D started in the end of 2019
• Participation in DeepFake Detection Challenge in Spring 2020
• Ranked among top 5% of solutions
• Alpha version internally released in Summer 2020
• Internally tested and evaluated by WeVerify partners in eight cycles
and continuously refined
• Version 1.0.0 released in November 2021
• Addition of network trained on more realistic datasets
• Available as a standalone service and via third party
applications: a) Truly Media, b) WeVerify plugin (soon)
20. Overview of Service
Input Images/Videos Pre-Processing Deep Learning Post-Processing Results / UI
- Shot segmentation
- 64 frames per shot
- Face detection
- Face filtering
- Face clustering per
shot and filtering
P. Charitidis, G. Kordopatis-Zilos, S. Papadopoulos and I. Kompatsiaris. “Investigating
the Impact of Pre-processing and Prediction Aggregation on the DeepFake Detection
Task”. In Proceedings of the Truth and Trust Online, 2020.
21. Overview of Service
Input Images/Videos Pre-Processing Deep Learning Post-Processing Results / UI
- Ensemble of models
(EfficientNet +
Transformers)
- Trained on DFDC
(120K videos) and
WildDeepFake (7314
videos) datasets
- BCE / InfoNCE loss
- DF scores per face
22. Overview of Service
Input Images/Videos Pre-Processing Deep Learning Post-Processing Results / UI
- Average DF scores
per face cluster
- Final prediction is
the maximum face
DF score
- Result preparation
24. Limitations in Detection
Hard to spot very
realistic manipulations
from methods that
involve manual tuning
and post-processing.
Current version cannot
detect manipulations in fully
synthetic faces (e.g.
StyleGAN2,
thispersondoesnotexist.com).
Low resolution faces may be falsely
presumed as DeepFakes.
25. Challenges
• Computational resources (both for training and for serving requests)
• Making the User Interface easy to understand
• Defend against adversarial approaches
• Generalization!
• Keeping up-to-date with new generation models/methods/tools
continuously enrich training dataset
26. Current Trends
Generate own DF and Use for
Training
Attention-based and Patch-
level Consistency Analysis
Metric and Contrastive
Learning
Domain Adaptation and
Knowledge Distillation
27. Next Steps
• New approaches
• Knowledge Distillation: from simple teacher-student pairs to group teaching setups
• Contrastive Learning: investigate decorrelated representations
• Practical considerations
• Usability
• Efficiency
• Maintenance (new training data, model adaptation, etc.)
• Transparency and Robustness
• Creating a model card for the service (modelcards.withgoogle.com)
• Benchmark service robustness with ART (github.com/Trusted-AI/adversarial-
robustness-toolbox)
28. Our DeepFake Detection Team
Akis: MeVer
leader
Nikos: MeVer senior
researcher
Panagiotis: service/API
development
Lazaros: front end
development
Spiros: DeepFake detection
and service development
Pantelis: GAN
detection
George: Deep Learning
research lead
Olga: technical
support
29. Thank you!
Dr. Symeon Papadopoulos
papadop@iti.gr
@sympap
Dr. Nikos Sarris
nsarris@iti.gr
@nikossarris
Media Verification (MeVer)
https://mever.iti.gr/
@meverteam