Can you prove where content or data came from, and that no one altered it? We build provenance systems that are hard to fake and easy to check, even when parts of the system are untrusted. Track A (AI-generated content): implement watermarking schemes whose keys are split among several parties, make every watermarking action publicly auditable, and use trusted execution environments to attest which model produced an output. Track B (sensor/IoT data): emulate a network of edge devices, attack its signed provenance records (replay, splicing, rollback, key misuse), and design verification at fog gateways that catches these attacks.
Why it matters. AI can now produce text, images and video that are hard to tell from human work. Military and infrastructure decisions also rely more and more on data from sensors placed where attackers can reach them. Both raise the same question: where did this come from, and has it been changed? Today’s answers usually require blind trust. An AI watermark is typically only as trustworthy as the single company holding its secret key, and anyone who steals that key can forge or strip watermarks. A signed sensor record helps only if replayed, spliced or rolled-back records get caught. We’re building provenance that holds up without blind trust.
Why our lab. The Data Security and Privacy Lab (DSP-Lab) works where security meets AI: homomorphic encryption, trusted hardware (Intel SGX, confidential VMs), blockchain-based data provenance and accountable AI. Undergraduates here are real team members:
You’ll get a structured onboarding package with a reading plan, hands-on exercises and self-check questions.
You’ll attend group meetings with Ph.D. students and work directly with me.
You’ll own a piece of a real system. E.g., our current undergraduates are building parts of a decentralized file-sharing platform and an encrypted large language model.