DeepFake Mini-Hackathon Wiki

Come join us for a fun and innovative hackathon exploring the DeepFake technology! Let's build exciting projects together 🎉

1. General Information
2. Themes
3. Resources
4. Schedule
5. Team Formation
6. Submission Process
7. Communication
8. Guidelines
9. Evaluation Criteria
10. Jury
11. Acknowledgements

1. General Information

Date & Time:

Date: Friday, September 13 to Saturday, September 14
Time: 6:00 PM - 10:00 PM (Sep 13) to 10:00 AM - 5:00 PM (Sep 14)

Location:

Those with "Approved" status on Luma: EPFL AI Center Lounge (ELE 117)
Everyone else: Zoom

2. Themes

LauzHack is hosting a mini-hackathon with two exciting tracks:

Generation (red team)
Detection (blue team)

3. Resources

Generation:

Face swap tools: https://analyticsindiamag.com/topics/best-face-swap-ai-tools/
Real-time face swap: https://github.com/hacksider/Deep-Live-Cam
List of generation resources: https://github.com/Daisy-Zhang/Awesome-Deepfakes
Hugging Face course on Diffusion models: https://www.deeplearning.ai/short-courses/open-source-models-hugging-face/
(Login required) Short courses from DeepLearning.ai (LangChain, RAG, Vector Databases, etc): https://learn.deeplearning.ai/
Actively developing Face Swapping framework to replace DeepFaceLab: https://github.com/deepfakes/faceswap

Deepfake datasets:

The most popular dataset of deepfake videos FaceForensics++
Another popular dataset of video deepfakes: Celeb-DF v2
Image detection challenge dataset: https://www.dfad.unimore.it/challenge/
Speaker verification and spoofing: https://www.asvspoof.org/
Partially spoofed audio dataset: https://zenodo.org/records/5766198
In-the-wild audio deepfakes from politicians and public figures: https://deepfake-total.com/in_the_wild, https://arxiv.org/pdf/2404.13914
Audio-Video Multimodal Deepfake Dataset FakeAVCeleb
Deepfake Detection Challenge Dataset from Facebook DFDC
Large video dataset of deepfakes by Idiap DF-Mobio]
Audio-video multimodal deepfake high quality dataset by Idiap SWAN-DF

Detection:

List of detection resources: https://github.com/Daisy-Zhang/Awesome-Deepfakes-Detection
For faces, a normal model (EfficientNet, Xception, ResNets) pretrained on faces should work fine for starters.
For voices, a model for speech segmentation could be a start such this by Pyannote on HuggingFace

Evaluation:

Compute metrics and ROC/DET curves using Scikit-learn
Metrics for segment-wise evaluations by Pyannote toolite developed for speaker diarization.

Compute / demo:

You can message Eric (organizer) on the LauzHack Discord for an OpenAI key!
Google Colab for GPU compute: https://colab.research.google.com/
Getting started with EPFL clusters: https://github.com/epfml/getting-started-lauzhack/
Gradio for hosting a demo.

4. Schedule

Friday, September 13 (on Zoom for those without the Luma registration approval):

6:00 PM: Tutorials/Workshops (AI Center Lounge)
- Talks/Tutorials, slides
- Pizza
- Spontaneous tutorials based on people's interests
10:00 PM: End of day. You CANNOT stay overnight, but you can continue working remotely.

Saturday, September 14 (only for participants whose Luma registration was approved):

10:00 AM: Breakfast (AI Center Lounge)
Until 3:00 PM: Hack, hack, hack!
3:30 PM: Demos (for everyone) then prizes 🏆

😋 We will provide dinner on Friday and (breakfast, lunch) on Saturday.

5. Team Formation

Up to 4 members per team. Register your team here.

6. Submission Process

⚠️ All projects should be submitted on DevPost. And make sure your team is up-to-date in the team formation document.

In order to be considered for a prize, all projects should demo/present (3 minutes) on Saturday afternoon.

When submitting your project, in the Additional info step, please select the track that you are doing (Generation or Detection) as in the attached image.

7. Communication

Real-time information about the event, food service details, and questions related to the challenges will be posted in our personal Discord server (#deepfake-sep2024 channel). Please use the link sent to you via Luma (to get access to the private channel for this event).

8. Guidelines

General Guidelines are provided by the LauzHack rules, though if you choose to be part of a red team, the following additional guidelines apply.

As deep fakes are a sensitive topic, and this hackathon explores both blue team (defensive) and red team (offensive) approaches, special rules apply. The following rule applies to everyone in attendance: Consider this a private event. Do not record any deep fakes created during the hackathon. If you have any question regarding this rule, ask the organizers first. This is very important as you might be liable for any harm caused by your actions.

The following additional rules apply to the Red team:

Generation track (Red Team) Rules

Our goal is not to limit your creativity - on the contrary, we encourage you to push the boundaries and show us what’s possible! However, it’s important that your well-intentioned work does not inadvertently cause harm.

As a principle, do not create deep fakes of real people
For public figures (famous people for which there is a lot of content online), you are allowed to create deep fakes in the scope of this event - see additional rules below
Every deep fake you create must be very clearly labeled as such
Refrain from producing deeply offensive content
No sharing or publishing of generated content or code outside the event without review
If in doubt: please regularly check-in with organizers to ensure adherence to guidelines

9. Evaluation Criteria

Only work done during hackathon will be considered (and should be made explicit) for the project evaluation.

Criteria	Allocated points
1. Technical Impressiveness	6
1.1. How impressive is the project from a technical perspective?	3
1.2. How reasonable the technical and programming solutions are, given the limited timeframe of a hackathon?	3
2. Idea	6
2.1. How innovative, original and unexpected the project is?	3
2.2. How usable the idea is for the real-world target population to which the project is aimed? hackathon?	3
3. Prototype	8
3.1. Does the prototype work as advertised by the team, and as expected for a one-day project?	4
3.2. Does the prototype provide a good user experience and usability?	4
4. Presentation	4
5. Integrity Check
TOTAL	24