Security in the age of AI

Adversarial Robustness

As AI systems are deployed in high-stakes applications, they become targets for adversaries. Our research investigates vulnerabilities in modern AI systems and develops defenses to make them more robust.

This includes work on prompt injection attacks against language models, model extraction and membership inference attacks, data poisoning during training, and federated learning security. We take a red-team approach, finding vulnerabilities ourselves so we can fix them before malicious actors exploit them.

Research Topics

Prompt injection detection and defenseModel extraction and intellectual property protectionData poisoning and backdoor detectionFederated learning security and privacyAdversarial example detection and filteringSecure multi-party computation for ML

Key Findings

Research in progress, findings coming soon

Future Directions

Formal verification of AI system security properties
Adversarial training that generalizes to novel attack types
Security benchmarks and red-teaming frameworks

Interested in this research area?

GET IN TOUCH