Organizing a Privacy-preserving Hackathon
Open-source for the win!
One of the core values shared by Zama, Entrepreneur First, and Hugging Face is our deep commitment to open-source. At all three companies, we not only advocate for open-source models but also develop our tools in the open. This allows people to explore, test, and prototype with them freely.
Hugging Face is known for its development of the Transformers library, while Zama is behind Concrete ML. Meanwhile, Entrepreneur First excels at selecting the right talent and helping them build groundbreaking projects and companies.
What Privacy-preserving AI, or Privacy-preserving ML
As a result, all three companies were thrilled to organize the first-ever Privacy-Preserving AI Hackathon.
Privacy-preserving machine learning focuses on delivering the same services—like inferences or training—but in a way that protects the user's privacy. Essentially, it allows third parties to analyze your data using ML without ever seeing your data in clear form. It sounds almost too magical to be true, but thanks to a cryptographic breakthrough called Fully Homomorphic Encryption (FHE), it's now possible. No more reliance on trust or worrying about security breaches—the security lies in the math.
At Zama, we used Hugging Face Spaces to demonstrate this technology, giving users (both companies and data scientists) the opportunity to see how it works and how it can be useful—without needing to dive into the technical details (and trust us, you don't want to get lost in the technical intricacies). For example, one project involved analyzing ancestries from encrypted DNA.
About the event
In September 2024, we hosted our event at Zama’s offices over two exciting days. Fifty participants, carefully selected by Entrepreneur First from over 200 applicants, joined us. Unfortunately, we couldn’t accommodate everyone due to overwhelming interest from the community. To those we couldn't invite, we sincerely apologize.
The hackathon offered two main challenges:
- Create a Hugging Face Space that demonstrates the capabilities of Concrete ML for a real-world business use case.
- Or, use Hugging Face endpoints to deploy a Concrete ML model and integrate it into an application.
Our panel of judges included four experts: Régis from Hugging Face, Guillaume from Entrepreneur First, Andrei from Zama, and Elad from Lunar Ventures, a venture capital firm investing in tech-founders with global ambition.
About the hackathon
Over the course of the two-day hackathon, teams formed, business ideas were brainstormed, and the final results were presented during the jury session. You can watch the video of the final contest here. The different Hugging Face Spaces created by participants can be found on the Hugging Face organization page of the event. Please note, since it was a fast-paced, two-day event, some spaces may not be fully functional yet, as many projects are still a work in progress.
We witnessed an incredible amount of creativity throughout the hackathon, with participants coming up with great business ideas. Applications ranged from health and sports, intrusion prevention, and audio analysis to patent research, identity verification, deepfake detection, and even model watermarking.
What impressed us the most was the high level of technical skill displayed, especially by those who were entirely new to Hugging Face Spaces, endpoints, privacy-preserving ML, or Concrete ML. It was a testament not only to the potential of this groundbreaking technology but also to how accessible our tools are for data scientists—whereas, historically, similar solutions have been restricted to the niche world of cryptographers.
About the winning projects
The jury selected three winners:
#3: Team Deep Fake Detection: This team introduced a deepfake detection system using FHE. Users can submit an encrypted image, and the service will determine if it's a deepfake—without ever seeing the image.
#2: Team Zamark: They developed a watermarking tool using FHE to generate model watermarks with an SGD classifier.
#1: Team Parseling: This team won the competition with a medical assistant based on an LLM, allowing additional analysis in FHE on untrusted services.
Feel free to explore these spaces—and because we’re open-source!—you can even check out the source code for these projects in the Files submenu (e.g., here). Take a look to see how Gradio and Concrete ML models were used, and what remains to be done!
Our Favorite Picks
To wrap up this blog post, we’d like to highlight a couple of our personal favorites beyond the previously mentioned winners.
For @binoua, it would be the two following spaces:
- scanning files for virus, without sending the files to the scan provider
- in-vehicule intrusion detection, to detect malware in a more and more complex environment
And for @regisss:
- accelerating FHE model inference using two models, one for non-sensitive tokens (no need for FHE) and one for sensitive tokens (requires FHE)
- privacy-preserving de-identification of audio files, to remove user-sensitive data from any recording
Now, it's your turn to build the future of privacy-preserving AI!
We had an amazing time during this hackathon, connecting with the community and helping participants build innovative products. Whether you were part of the event and want to continue, or you’re new and looking to get involved, don’t hesitate to:
- Join the Zama community channels to ask your technical questions.
- Join the Hugging Face Discord to engage with the broader HF community.
- Apply to Entrepreneur First programs and kickstart your journey as a founder.
Feel free to ping us with your best Hugging Face Spaces—we’d love to check them out and help promote them!
