Machine Learning Security Evasion Competition 2020
summary
A competition modifies 50 Windows malware samples to evade three ML engines (EMBER plus two submitted models), first as a white‑box then as a black‑box challenge; 70 people register in early rounds, 11 bypass at least one model, later 60 register and five bypass any model, engines see ~5,000 queries. Practical wins use simple file edits: the overlay technique (appending data) is the most effective, others use added sections borrowed from benign files or packers that often break functionality; example tooling includes EMBER, hyperopt, VMRay and Docker defenses.
AI-written summary · about
chapters (5)
- 0:00Start
- 13:04Demo
- 17:07Winners of 2020
- 19:11Machine Learning Evasion Code Walkthrough
- 20:15Leaderboard Criteria
related talks
53mSecurity of Machine Learning Systems
28mSecurity of Machine Learning
52mSleeping With One AI Open: An Introduction to Attacks Against...
33mProactive EDR Against Adaptive Evasion: Countering Self‑Modifying Agentic Malware
36mAdversarial Machine Learning: How Secure Is ML?
20mMachine Learning Attacks – A New Era of Security Epidemic
26mWhen Models Write Malware: Defending Against On-demand Exploits
49mDo Attackers Use Algorithms to Evade ML? An In-the-Wild Phishing Case Study