Saved in:
Bibliographic Details
Main Authors: Xiong, Chen, Wang, Zihao, Zhu, Rui, Ho, Tsung-Yi, Chen, Pin-Yu, Xiong, Jingwei, Tang, Haixu
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2506.06057
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914526700503040
author Xiong, Chen
Wang, Zihao
Zhu, Rui
Ho, Tsung-Yi
Chen, Pin-Yu
Xiong, Jingwei
Tang, Haixu
author_facet Xiong, Chen
Wang, Zihao
Zhu, Rui
Ho, Tsung-Yi
Chen, Pin-Yu
Xiong, Jingwei
Tang, Haixu
contents Large Language Models (LLMs) rely on massive training datasets, often including proprietary data, which raises concerns about unauthorized usage and copyright infringement. Existing dataset inference methods typically require access to log probabilities or other internal signals, but many modern LLMs restrict such access, motivating token-only inference approaches. We propose CatShift, a token-only dataset inference framework based on catastrophic forgetting, where models overwrite prior knowledge when trained on new data. Fine-tuning an LLM on a subset of its training data induces larger output shifts than fine-tuning on unseen data. CatShift compares these shifts against those from a known non-member validation set to infer whether a dataset was included in training. Experiments on both open-source and API-based LLMs show that CatShift remains effective without logit access, enabling practical protection of proprietary datasets.
format Preprint
id arxiv_https___arxiv_org_abs_2506_06057
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Hey, That's My Data! Token-Only Dataset Inference in Large Language Models
Xiong, Chen
Wang, Zihao
Zhu, Rui
Ho, Tsung-Yi
Chen, Pin-Yu
Xiong, Jingwei
Tang, Haixu
Computation and Language
Artificial Intelligence
Large Language Models (LLMs) rely on massive training datasets, often including proprietary data, which raises concerns about unauthorized usage and copyright infringement. Existing dataset inference methods typically require access to log probabilities or other internal signals, but many modern LLMs restrict such access, motivating token-only inference approaches. We propose CatShift, a token-only dataset inference framework based on catastrophic forgetting, where models overwrite prior knowledge when trained on new data. Fine-tuning an LLM on a subset of its training data induces larger output shifts than fine-tuning on unseen data. CatShift compares these shifts against those from a known non-member validation set to infer whether a dataset was included in training. Experiments on both open-source and API-based LLMs show that CatShift remains effective without logit access, enabling practical protection of proprietary datasets.
title Hey, That's My Data! Token-Only Dataset Inference in Large Language Models
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2506.06057