Competitions in Infosec ML

### The Machine Learning Security Evasion Competition At 00:00, Dr. Hyrum Anderson (Endgame) and Ryan Kovar (Splunk) announce the Machine Learning Security Evasion Competition (MLSEC), inviting researchers to participate in testing machine learning malware classifiers.

### The Adversarial Challenge: Functional Malware Mutation At 02:02, Anderson outlines the core attacker objective: competitors must modify real Windows PE malware samples to bypass machine learning detection models while preserving the binary's malicious runtime execution.

### Automated Evaluation Harness and Sanity Checking At 04:01 and 06:01, the presenters explain the automated scoring architecture. Submitted binaries pass through rigorous verification sandboxes to ensure file integrity and confirm that core functionality (such as ransomware file encryption or command-and-control beacons) remains operational.

### Attacking Ember and Non-Negative Weight Models At 08:00, Anderson analyzes the targeted defensive models. Contestants compete against the open-source Ember classifier and specialized deep neural networks trained with non-negative weight constraints, designed to resist adversarial feature dilution.

### Black-Box Attack Strategies and Guidance Concluding between 10:01 and 12:01, the speakers share optimization strategies for contestants. They recommend leveraging reinforcement learning agents, genetic algorithms, and local surrogate models to discover minimal byte mutations that successfully evade commercial ML endpoints.

(AI summary — Gemini Flash)