arXiv 论文:LLM "指纹"会传染,API 中转审计方法需重写
Contagious Fingerprints: When LLMs Inherit Another Model's Identity
做 API 审计或依赖第三方中转的开发者要看:论文证明换模型后指纹会跟着对话历史走,旧检测方法直接失效,DeepSeek 替 Claude Sonnet 检出率掉到 0.72。新方法 LINEUP 不需要阈值就能在 4 轮内抓到 0.81 的顶替会话。
arXiv 论文(编号 2611.18104)发现 LLM 的文本"指纹"具有传染性:模型接手另一模型开启的对话后,会继承前者的指纹。实验显示,DeepSeek 顶替 Claude Sonnet 后,文本分类器的检出率从 0.97 降至 0.72;同厂商降级(Haiku 顶替)时从 0.87 降至 0.54。这意味着 API 中转商可以先用客户付费的模型服务前几轮对话,再悄悄切给廉价模型,而现有审计难以发现。
传染的机制被归因于对历史对话风格的模仿。论文还指出,基于新鲜对话校准的主动探针在长上下文中对诚实模型的误报率最高达 1.00,标准 MMD 等同性检验在全部 9 个测试模型上都判定异常——根本原因是把"依赖历史的回复"与"无历史的参照"做比较。
作者提出 LINEUP 方法:从被声称的模型在同一对话历史上采样参照回复,按排名检验被审计回复。该方法通过可交换性控制误报,无需拟合阈值或对数概率。在两个闭源模型上的 contemporaneous 参照下,误报率为 0.056(预注册上限 0.10),4 轮审计后对 6 组模型配对的顶替会话检出率达 0.81。
Contagious Fingerprints: When LLMs Inherit Another Model's Identity
The fingerprint by which a classifier recognizes the LLM behind a text is usually treated as a fixed property of the model. We show that it is contagious: a model that continues a conversation begun by another model inherits that model's fingerprint. This undermines audits of API relays, which can serve the first turns of a conversation with the model a customer pays for and silently hand the rest to a cheaper one. After such a handoff, detection of the substitute by a text classifier drops from 0.97 to 0.72 (DeepSeek replacing Claude Sonnet) and from 0.87 to 0.54 for a same-vendor downgrade (Haiku), and retraining on handoff data repairs only substitutes seen in training. Preregistered experiments trace the contagion to imitation of the history's style. History also moves a model away from its own fingerprint: active probes are not camouflaged, but when calibrated on fresh conversations they flag \emph{honest} models on long contexts at rates up to 1.00, and a standard MMD equality test rejects all nine models we test. Both failures come from comparing a history-dependent reply with a history-free reference. We propose \textsc{Lineup}, which samples references from the claimed model \emph{on the same history} and tests the audited reply by its rank among them. \textsc{Lineup} controls false positives by exchangeability and needs no fitted threshold or log-probabilities. With contemporaneous references on two closed models, its false-positive rate is 0.056 (preregistered bound 0.10), and it detects 0.81 of substitute sessions across six pairs after four audited turns.
This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include arXiv, LINEUP.
Primary source verification: published directly by arXiv: DeepSeek with first-party authority.