TMI! Finetuned Models Leak Private Information from their Pretraining Data

TL;DR

A novel metaclassifier-based attack that leverages the influence of memorized pretraining samples on predictions in the downstream task, TMI, is implemented and it is found that TMI can successfully infer membership of pretraining examples using query access to the finetuned model.

Abstract

Transfer learning has become an increasingly popular technique in machine learning as a way to leverage a pretrained model trained for one task to assist with building a finetuned model for a related task. This paradigm has been especially popular for $\textit{privacy}$ in machine learning, where the pretrained model is considered public, and only the data for finetuning is considered sensitive. However, there are reasons to believe that the data used for pretraining is still sensitive, making it essential to understand how much information the finetuned model leaks about the pretraining data. In this work we propose a new membership-inference threat model where the adversary only has access to the finetuned model and would like to infer the membership of the pretraining data. To realize this threat model, we implement a novel metaclassifier-based attack, $\textbf{TMI}$, that leverages the influence of memorized pretraining samples on predictions in the downstream task. We evaluate $\textbf{TMI}$ on both vision and natural language tasks across multiple transfer learning settings, including finetuning with differential privacy. Through our evaluation, we find that $\textbf{TMI}$ can successfully infer membership of pretraining examples using query access to the finetuned model. An open-source implementation of $\textbf{TMI}$ can be found on GitHub: https://github.com/johnmath/tmi-pets24.

Figures (17)

• Figure 1: Our New Membership-Inference Threat Model.
• Figure 2: Distribution of the Test Statistic, $z$, for Multiple Values of $\alpha$
• Figure 3: AUC of our Membership-Inference Attack on Mean Estimation as a Function of $\alpha$.
• Figure 4: Scaled Model Confidences of Shadow Models Finetuned on Caltech 101 at Multiple Labels when Queried on a Sample from the "Dugong" Class in Tiny Imagenet
• Figure 5: TMI Attack Performance on Downstream Tasks When Preterained CIFAR-100 Target Models are Finetuned Using Feature Extraction

Theorems & Definitions (11)

• Example 1.1
• Example 1.2
• Definition 2.1
• Lemma 4.1
• Lemma 4.2
• Lemma A.1
• proof
• Lemma A.2
• proof
• Lemma A.3