← All talks

Cybersecurity's Data Labeling Problem: Reinforcement Learning Solution #shorts

BSides Frankfurt1:084 viewsPublished 2025-12Watch on YouTube ↗
About this talk
Cybersecurity faces a major hurdle: limited labeled data. The complexity is immense, unlike simple homework problems. How can reinforcement learning unlock creative solutions and improve models? Good oracles and feedback are key. #Cybersecurity #DataScience #MachineLearning #ReinforcementLearning
Show transcript [en]

What's important is in the past in many cases in cyber security we've tried really to label the data and that's our biggest problem. We don't have a lot of labelled data because the complexity of cyber security is just so bad. See this as homework you're giving to your your children and then ask them solve this homework and this is the answer to that. It's not going to make smart children out of that right but playing games again with a scoreboard where they can improve all the time on a daily basis and become better Counter Strike players. That's something that they they they need and indeed we see a lot of research day by day in this environment

where for instance recently there was a very nice research where um uh where just with a single function and a self-training selfplay against each other a model became better than any other model in in mass just by ensuring that we have really good oracles and really good feedback system to the specific function when this mass models were built. So reinforcement learning can unlock a lot of this creativity if we can um if we can create the right problems.