Data-Free Privacy-Preserving for LLMs via Model Inversion and Selective Unlearning

TL;DR

This paper introduces a data-free method for removing sensitive PII from large language models by synthesizing pseudo-PII, creating privacy masks, and performing token-level unlearning, thus enhancing privacy without access to original training data.

Contribution

The paper presents a novel data-free framework called Data-Free Selective Unlearning (DFSU) that removes PII from LLMs without needing training data, using model inversion and contrastive mask loss.

Findings

Effectively removes target PII from LLMs

Maintains model utility after unlearning

Demonstrates success on Pythia models and PII-Masking dataset

Abstract

Large language models (LLMs) exhibit powerful capabilities but risk memorizing sensitive personally identifiable information (PII) from their training data, posing significant privacy concerns. While machine unlearning techniques aim to remove such data, they predominantly depend on access to the training data. This requirement is often impractical, as training data in real-world deployments is commonly proprietary or inaccessible. To address this limitation, we propose Data-Free Selective Unlearning (DFSU), a novel privacy-preserving framework that removes sensitive PII from an LLM without requiring its training data. Our approach first synthesizes pseudo-PII through language model inversion, then constructs token-level privacy masks for these synthetic samples, and finally performs token-level selective unlearning via a contrastive mask loss within a low-rank adaptation (LoRA) subspace. Extensive experiments on the AI4Privacy PII-Masking dataset using Pythia models demonstrate that our method effectively removes target PII while maintaining model utility.