Scaling behavior of LLM-based deanonymization to internet-scale candidate pools

Prove that the large-language-model-based deanonymization pipeline that extracts micro-data from user comments as structured summaries, searches via cosine similarity over dense text embeddings, selects from the top-k candidates using an LLM, and calibrates match decisions via pairwise LLM comparisons maintains non-trivial recall at high precision (e.g., 90% precision) when the candidate pool size grows to internet-scale (on the order of one million users), consistent with the observed log-linear scaling from smaller pools.

Background

In the temporal-split Reddit experiment, the authors vary the candidate pool size and observe that recall at 90% precision for their strongest LLM-based attack (Search + Reason + Calibrate) degrades roughly log-linearly with pool size. Extrapolation from empirical results up to 10,000 candidates projects about 35% recall at 1,000,000 candidates.

Because this projection is based on a coarse extrapolation, the authors explicitly conjecture that LLM deanonymization will retain non-trivial success at internet-scale, posing a significant privacy threat if validated. Formal confirmation would require establishing the scaling law and demonstrating sustained high-precision recall at much larger pool sizes.