Trustworthy Visual Privacy Auditing with Causal Governance and Resilient Federated Protection for NIST AI Risk Management Framework

Our previous visual privacy framework leveraging Graph Convolutional Networks (GCNs) and Federated Learning (FL) has been shown to achieve state-of-the-art (SOTA) predictive performance. However, it neglects the systemic requirements of the National Institute of Standards and Technology (NIST) AI Risk Management Framework (AI RMF). To address this critical gap, this paper proposes the Trustworthy Visual Privacy Auditing (TVPA) system, which transitions conventional static detection models into a dynamic and secure governance ecosystem. We first establish system resilience against adversarial threats by proposing an active auditing mechanism called Resilient Federated Protection (RFP) to embed unique model parameter watermarks within client-side updates. The RFP mechanism enables the federated aggregator to verify node legitimacy and automatically isolate malicious clients attempting poisoning attacks. Then, to ensure strict accountability, we design an immutable audit log mechanism in the RFP mechanism that utilizes a Cryptographic Hash Chain (CHC) to record and verify the provenance of every model update, creating a transparent chain of custody. Furthermore, the prediction mechanism is enhanced by Causal Governance (CG) that integrates causal inference to provide counterfactual reasoning for explaining the root causes of privacy risks rather than merely flagging associations. Experiments on the VISPR dataset demonstrate that our TVPA system can synthesize high-performance recognition with robust security, auditability, and causal explainability to provide trustworthy AI governance.