Maksim and his team faced an ambitious challenge: detecting malicious code using machine learning, without any labeled data. Labeling such data requires deep domain expertise and a lot of expert time. To overcome this, we automated the labeling process using LLMs and successfully solved the initial problem.
In this talk, the speaker will share his experience building this pipeline:
- how using LLMs he managed to speed the labeling process up a zillion times
- how he grew the team’s malicious sample set from 373 to over 6,500 files
The speaker will also compare LLMs from different vendors and explain why, despite all this, AI still can’t replace cybersecurity experts.
This session will be valuable for AI enthusiasts, cybersecurity experts, and ML engineers working in this field
SPEAKERS
Maksim Mitrofanov
ML Team Lead, Application Security Analysis, Positive Technologies
Maksim leads the Machine Learning Group for Application Security at Positive Technologies. He focuses on developing ML/LLM systems that improve cybersecurity workflows and power core products such as PT Application Firewall, PT Data Security, and PT Application Inspector.
His recent projects include malicious code analysis, sensitive data classification, and machine learning systems for DDoS protection
