GraphProfiler addresses the privacy risks associated with LLM-based attribute inference. The system constructs a personal knowledge graph for each user, representing their post history. Nodes in the graph correspond to individual posts, and edges trace relationships between them. The system resolves attribute predictions to cited graph records and the original source texts. The research achieved 86.7% attack success rate on the SynthPAI benchmark and 84.6% on the PANDORA benchmark. These results are within two points of strong text-only baselines. The system cited supporting evidence for over 98% of predictions. Ablation experiments demonstrated that removing the cited posts significantly reduced the attack success rate compared to removing random posts. This indicates the importance of source attribution in mitigating inference attacks. The GraphProfiler system provides a mechanism for understanding the specific posts that contribute to sensitive attribute inference, enabling focused redaction or rewriting strategies. The system’s architecture allows for targeted privacy mitigation, rather than broad perturbation of user histories.
Source: https://arxiv.org/abs/2609.12448