Software Engineering Institute | Carnegie Mellon University reposted this
Recorded a podcast episode with Julie Lawler for the SEI Podcast Series on data poisoning in AI systems and the case for cryptographic chain of custody as a defense. The conversation draws on recent work from our colleagues Renae Metcalf and Matthew Churilla. We covered what data poisoning is and why training-time attacks are hard to diagnose, where current mitigations fall short, and how chain of custody (using cryptographic hashes, digital signatures, and append-only logs) provides a way to make tampering detectable across the data pipeline. We worked through a drone-based example, talked about where a team should pragmatically start if they're building this from scratch, and closed on how the approach maps onto zero trust principles. If you're working on safety-critical AI in industry, academia, or government and this kind of pipeline integrity work is relevant to what you're doing, the SEI is open to collaborations — info@sei.cmu.edu is the right contact. https://lnkd.in/gxgHqWgA