ReCaLL: Membership Inference via Relative Conditional Log-Likelihoods

TL;DR

ReCaLL is a new membership inference attack that detects whether data was used in training large language models by analyzing changes in conditional log-likelihoods, achieving state-of-the-art results.

Contribution

The paper introduces ReCaLL, a novel method leveraging relative conditional log-likelihoods for effective membership inference on LLMs, with comprehensive experimental validation.

Findings

ReCaLL outperforms existing MIAs on WikiMIA dataset.

Conditioning on non-member prefixes causes larger log-likelihood decreases for members.

Ensemble approaches further improve ReCaLL's inference accuracy.

Abstract

The rapid scaling of large language models (LLMs) has raised concerns about the transparency and fair use of the data used in their pretraining. Detecting such content is challenging due to the scale of the data and limited exposure of each instance during training. We propose ReCaLL (Relative Conditional Log-Likelihood), a novel membership inference attack (MIA) to detect LLMs' pretraining data by leveraging their conditional language modeling capabilities. ReCaLL examines the relative change in conditional log-likelihoods when prefixing target data points with non-member context. Our empirical findings show that conditioning member data on non-member prefixes induces a larger decrease in log-likelihood compared to non-member data. We conduct comprehensive experiments and show that ReCaLL achieves state-of-the-art performance on the WikiMIA dataset, even with random and synthetic prefixes, and can be further improved using an ensemble approach. Moreover, we conduct an in-depth analysis of LLMs' behavior with different membership contexts, providing insights into how LLMs leverage membership information for effective inference at both the sequence and token level.