STALKER: Composed Quasi-Identifiers and LLM-Amplified Behavioral Leakage in Anonymized Marketplace Data
Year: 2026
Type: conference-paper
Abstract: Online resale marketplaces routinely apply basic anonymization techniques, including pseudonymization of user identifiers and normalization of event timestamps, to mitigate privacy risks prior to data release. This paper provides empirical evidence that such measures are insufficient for protecting behavioral privacy in consumer-to-consumer (C2C) platforms. Using two subsets drawn from a large-scale dataset, MerRec, from the Mercari marketplace, we show that the composition of structured quasi-identifiers, stable temporal activity patterns, and semantic signals embedded in free-text item titles enables the reconstruction of detailed and distinctive behavioral profiles without access to external data or real-world identity information. Behavioral analysis examines 92,034 anonymized transaction records from the ten highest-activity users, while population-scale leakage is quantified over a random sample of 500 users of 1-16,275 events.
Institution: Florida Polytechnic University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7164559633