Igor Shilov

Whether LLMs memorize their training data and what this means, from privacy leakage to detecting copyright violations – has become a rapidly growing area of research over the last two years. In recent months, more than 10 new methods have been proposed to perform Membership Inference Attacks (MIAs) against LLMs. Contrary to traditional MIAs which rely on fixed – but randomized – records or models, these methods are mostly evaluated on datasets collected post-hoc. Sets of members and non-members, used to evaluate the MIA, are constructed using informed guesses after the release of a model. This lack of randomization raises concerns of a distribution shift between members and non-members. In the first part, we review the literature on MIAs against LLMs. While most work focuses on sequence-level MIAs evaluated in post-hoc setups, we show that a range of target models, motivations and units of interest have been considered in the literature. We then quantify distribution shifts present in the 6 datasets used in the literature, ranging from books to papers, using a bag of word classifier. Our analysis reveals that all of them suffer from severe distribution shifts. This challenges the validity of using such setups to measure LLM memorization and may undermine the benchmarking of recently proposed methods. Yet, all hope might not be lost. In the second part, we introduce important considerations to properly evaluate MIAs against LLMs and discuss potential ways forward: randomized test splits, injections of randomized (unique) sequences, randomized finetuning, and post-hoc control methods. While each option comes with its advantages and limitations, we believe they collectively provide solid grounds to guide the development of MIA methods and study LLM memorization. We conclude by proposing comprehensive, easy-to-use benchmarks for sequence- and document-level MIAs against LLMs.

Questions of fair use of copyright-protected content to train Large Language Models (LLMs) are being very actively debated. Document-level inference has been proposed as a new task: inferring from black-box access to the trained model whether a piece of content has been seen during training. SOTA methods however rely on naturally occurring memorization of (part of) the content. While very effective against models that memorize a lot, we hypothesize–and later confirm–that they will not work against models that do not naturally memorize, e.g. medium-size 1B models. We here propose to use copyright traps, the inclusion of fictitious entries in original content, to detect the use of copyrighted materials in LLMs with a focus on models where memorization does not naturally occur. We carefully design an experimental setup, randomly inserting traps into original content (books) and train a 1.3B LLM. We first validate that the use of content in our target model would be undetectable using existing methods. We then show, contrary to intuition, that even medium-length trap sentences repeated a significant number of times (100) are not detectable using existing methods. However, we show that longer sequences repeated a large number of times can be reliably detected (AUC=0.75) and used as copyright traps. We further improve these results by studying how the number of times a sequence is seen improves detectability, how sequences with higher perplexity tend to be memorized more, and how taking context into account further improves detectability.

We consider the privacy-preserving machine learning (ML) setting where the trained model must satisfy differential privacy (DP) with respect to the labels of the training examples. We propose two novel approaches based on, respectively, the Laplace mechanism and the PATE framework, and demonstrate their effectiveness on standard benchmarks. While recent work by Ghazi et al. proposed Label DP schemes based on a randomized response mechanism, we argue that additive Laplace noise coupled with Bayesian inference (ALIBI) is a better fit for typical ML tasks. Moreover, we show how to achieve very strong privacy levels in some regimes, with our adaptation of the PATE framework that builds on recent advances in semi-supervised learning. We complement theoretical analysis of our algorithms’ privacy guarantees with empirical evaluation of their memorization properties. Our evaluation suggests that comparing different algorithms according to their provable DP guarantees can be misleading and favor a less private algorithm with a tighter analysis.

We introduce Opacus, a free, open-source PyTorch library for training deep learning models with differential privacy (hosted at opacus.ai). Opacus is designed for simplicity, flexibility, and speed. It provides a simple and user-friendly API, and enables machine learning practitioners to make a training pipeline private by adding as little as two lines to their code. It supports a wide variety of layers, including multi-head attention, convolution, LSTM, GRU (and generic RNN), and embedding, right out of the box and provides the means for supporting other user-defined layers. Opacus computes batched per-sample gradients, providing higher efficiency compared to the traditional "micro batch" approach. In this paper we present Opacus, detail the principles that drove its implementation and unique features, and benchmark it against other frameworks for training models with differential privacy as well as standard PyTorch.