Abstract:Recent research shows that large language models are susceptible to privacy attacks that infer aspects of the training data. However, it is unclear if simpler generative models, like topic models, share similar vulnerabilities. In this work, we propose an attack against topic models that can confidently identify members of the training data in Latent Dirichlet Allocation. Our results suggest that the privacy risks associated with generative modeling are not restricted to large neural models. Additionally, to mitigate these vulnerabilities, we explore differentially private (DP) topic modeling. We propose a framework for private topic modeling that incorporates DP vocabulary selection as a pre-processing step, and show that it improves privacy while having limited effects on practical utility.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to explore whether simpler generative models (such as topic models, especially Latent Dirichlet Allocation (LDA)) also have privacy risks, similar to the privacy attacks faced by large - scale language models. Specifically, the authors focus on whether these models can make Membership Inference Attacks (MIAs) possible due to the "memory" of the training data. ### Main Research Questions 1. **Privacy Vulnerability of Topic Models**: - The authors propose an attack method against LDA, which can confidently identify which documents are part of the training data. - This shows that privacy risks in generative modeling are not limited to large neural models, and even simpler topic models may face similar threats. 2. **Methods to Mitigate Privacy Vulnerabilities**: - To address these privacy vulnerabilities, the authors explore the application of Differential Privacy (DP) in topic modeling. - A framework is proposed, using DP vocabulary selection as a pre - processing step to enhance privacy protection while minimizing the negative impact on practicality. ### Specific Research Contents - **Membership Inference Attacks (MIAs)**: By designing query statistics specific to LDA and integrating them into the LiRA framework, it is shown how to use the memory characteristics of the model for effective membership inference attacks. - **Differential Privacy (DP) Topic Modeling**: A complete DP topic - modeling process is proposed, including DP vocabulary selection and learning methods, to ensure that no sensitive information is leaked when releasing the topic - word distribution. ### Mathematical Formulas - **Topic Model Definition**: \[ \text{For a vocabulary size of } V \text{ and } k \text{ topics, the topic model } \Phi \in [0,1]^{k\times V} \text{ is a matrix whose row sums are 1, representing the word distribution of each topic.} \] - **Likelihood Ratio Test Statistic**: \[ \Lambda(\Phi, d)=\frac{p(\Phi_{\text{obs}}|T_{\text{in}}(d))}{p(\Phi_{\text{obs}}|T_{\text{out}}(d))} \] where \(p(\Phi_{\text{obs}}|T_{x}(d))\) is the probability density function of observing \(\Phi_{\text{obs}}\) under \(T_{x}(d)\). - **Query Statistic**: \[ \zeta(\Phi, d)=\max_{\theta}\log p(d|\theta,\Phi)=\max_{\theta}\sum_{w\in d}\log\left(\sum_{z}\theta_{z}\cdot\Phi_{z,w}\right) \] where \(\theta\in[0,1]^{k}\) and \(\sum_{z}\theta_{z} = 1\). ### Conclusion This paper experimentally proves that even simple topic models (such as LDA) are vulnerable to privacy attacks, especially membership inference attacks. In addition, the authors propose a topic - modeling framework based on differential privacy, which can maintain a certain level of practicality while ensuring privacy. This research result is of great significance for understanding the privacy risks of generative models and developing more secure machine - learning models.

Membership Inference Attacks and Privacy in Topic Modeling

GAN-Leaks: A Taxonomy of Membership Inference Attacks against Generative Models

Membership Inference Attack Susceptibility of Clinical Language Models

Privacy-preserving Generative Framework Against Membership Inference Attacks

Investigating Membership Inference Attacks under Data Dependencies

SoK: Reducing the Vulnerability of Fine-tuned Language Models to Membership Inference Attacks

A Method to Facilitate Membership Inference Attacks in Deep Learning Models

Membership Inference Attacks Against In-Context Learning

Improved Membership Inference Attacks Against Language Classification Models

Private Data Inference Attacks against Cloud: Model, Technologies, and Research Directions

Privacy Risks of Securing Machine Learning Models against Adversarial Examples

Machine Learning with Membership Privacy using Adversarial Regularization

Generated Distributions Are All You Need for Membership Inference Attacks Against Generative Models

Noisy Neighbors: Efficient membership inference attacks against LLMs

Differentially Private Data Generative Models

Subject Membership Inference Attacks in Federated Learning

An Empirical Study on the Membership Inference Attack against Tabular Data Synthesis Models

Membership Inference Attacks against Language Models via Neighbourhood Comparison

Monte Carlo and Reconstruction Membership Inference Attacks against Generative Models

Systematic Evaluation of Privacy Risks of Machine Learning Models

Defenses to Membership Inference Attacks: A Survey