Hardware Challenge: Detect Stealth AI Training Runs

Join us for an innovative hardware security hackathon focused on building new methods to detect and control for secret training of AI models! Work with cutting-edge side-channel monitoring equipment, collaborate with experts in hardware security, and help develop crucial tools for AI governance. $2000 in prizes available across builder and breaker tracks.

• ✨ Access state-of-the-art specialized hardware setups for AI compute verification and monitoring experimentation
• 🔍 Develop detection algorithms on a new dataset of GPU side-channel monitoring (power usage, internet communication, etc.)
• 🛠️ Test against real training runs, such as NanoGPT inference on an H100 compute box

Overview

This hackathon brings together hardware experts, security researchers, and AI safety enthusiasts to develop and test verification systems for detecting AI training activities through side-channel analysis. Participants will work with state-of-the-art hardware setups to create innovative solutions for monitoring compute usage.

Why Participate?

• Access to specialized hardware including side-channel monitoring equipment
• Work with leading experts in hardware security and AI governance
• Contribute to important tools for AI training verification
• Network with professionals from major cloud providers and hardware manufacturers

Challenge Tracks

Builder Track

In the Builder track, you will create new methods to detect stealth AI training runs and new verification protocols to ensure responsible use of compute.

• Develop side-channel monitoring systems
• Create classification algorithms for training detection
• Build verification protocols for hardware governance
• Implement real-time monitoring solutions

Breaker Track

In the Breaker track, you'll be tasked to evade classification algorithms with new types of training, making it more difficult for the Builders (and at the same time showing us how easy it might be to evade existing verification methods).

• Test evasion techniques against monitoring systems
• Develop methods to mask training signatures
• Analyze system vulnerabilities
• Create adversarial examples