How I built the world’s most efficient deepfake detector with $100

At a time when disinformation and manipulation of opinion spread through social networks, it is necessary to be equipped to spot fake identities, especially via their profile pictures which are the hardest elements to fake. During this talk, you will discover a method to detect images generated via ultra-realistic deepfake generator, with 100% reliability and even a way to know the exact time the photo was produced.

The talk will focus on a technical challenge faced during the implementation of one of the attacks, which involves TensorFlow models leveraging features of a customized ElasticSearch instance, on very limited hardware.


  • Basics on Machine Learning is a plus but not necessary


  • How GANs work and how they are trained
  • How to build a very efficient facial indexing engine using FaceNet and Elasticsearch for k-NN search on vectors
  • How to build a security threat model around bots and fake social media profiles



Mathis Hammel is a tech evangelist at CodinGame, a website specialized in mini-games to learn programming. He is a specialist and technical advisor in cybersecurity, machine learning, and algorithms. Mathis is passionate about technical challenges such as programming competitions and Capture The Flag, and holds several titles from national and international championships.


Sie möchten über die data2day
auf dem Laufenden gehalten werden?