
Nathan Liénard, Lama Sleem, Jérôme François, Léo Lavaur, Lujun Li · HAL 2026 · 2026
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
The widespread deployment of Large Language Models (LLMs) through black-box APIs raises significant trust and compliance concerns, as users lack guarantees that the responding model matches the one advertised. This opacity enables risks such as silent model substitution for cost reduction or Man-in-the-Middle (MitM) attacks that redirect queries to malicious models. This paper presents MODELMATCH, a passive fingerprinting framework for verifying LLM identity through output-only analysis, without access to internal parameters or logits. The method leverages 69 stylistic and interaction-based features within a hierarchical classification architecture that performs family-level identification, resolution of conflicting identification, and intra-family classification.
No comments yet — start the discussion below.