πCreds: Privacy-Inferring Credentials

πCreds is a privacy-preserving, legacy-compatible decentralized verifiable credential generated by a trusted LLM reasoning over authentication data. Compared with existing credential systems, the LLM's semantic reasoning over unstructured data significantly expands the range of attestable claims; the authors formalize two new threats: SCAE (an attacker manipulates authentication data to obtain misleading credentials) and ACPP (privacy leakage through adversarial model selection). The prototype supports issuing credentials for real-time financial, health, email, and code data sources, and the authors empirically studied these two threats on product expertise credentials over real financial data.