# Membership Inference Attacks on Machine Learning: A Survey

**Authors:** Hongsheng Hu, Zoran Salcic, Lichao Sun, Gillian Dobbie, Philip S. Yu, Xuyun Zhang  
**arXiv:** [2103.07853](https://arxiv.org/abs/2103.07853)  
**Date:** 2021-03-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2103.07853v4/

## Key points

- Machine learning (ML) models have become popular and widely used in various applications
- Concerns have emerged regarding the vulnerability of ML models to membership inference attacks (MIAs)
- MIAs aim to determine if a specific data record was used to train a target model
- Successful MIAs can lead to privacy breaches and pose risks to individuals' privacy and confidentiality
- Multiple studies have shown that MIAs can be effective on different types of ML models
- Researchers have proposed various defense methods to mitigate these attacks
- No systematic survey has been conducted on membership inference attacks and defenses until now
- The authors present the first comprehensive survey on membership inference attacks and defenses in machine learning
- They provide taxonomies for both attacks and defenses based on their characterizations
- The paper identifies limitations, gaps, and future research directions in this field
- The survey serves as a valuable reference for the research community and provides clear explanations for researchers outside of this domain
- An online resource repository has been created by the authors for staying updated with relevant work in this area

## Summary

In recent years, machine learning (ML) models have become increasingly popular and widely used in various applications such as image classification, text generation, audio recognition, and graph data analysis. However, a concerning issue has emerged regarding the vulnerability of ML models to membership inference attacks (MIAs). These attacks aim to determine whether a specific data record was used to train a target model or not. The implications of successful MIAs are significant as they can directly lead to privacy breaches. For example, an attacker could infer that the owner of a clinical record has a certain disease by identifying if the record was used to train a model associated with that disease. This poses serious risks to individuals' privacy and confidentiality. Multiple studies have shown that MIAs can be effective on different types of ML models, including classification models and generative models. As a result, researchers have proposed various defense methods to mitigate these attacks. However, despite the growing importance of this research area, there has been no systematic survey conducted on membership inference attacks and defenses. To address this gap in knowledge, the authors of this paper present the first comprehensive survey on membership inference attacks and defenses in machine learning. They provide taxonomies for both attacks and defenses based on their characterizations and discuss their respective advantages and disadvantages. By identifying limitations and gaps in existing research, the authors also highlight several promising future directions for researchers interested in this field. The survey serves as a valuable reference for the research community while also providing clear explanations for researchers outside of this specific domain. To further assist researchers in staying updated with relevant work in this area, the authors have created an online resource repository which will be regularly updated with new findings. Interested readers can access this repository at https://github.com/HongshengHu/membership-inference-machine-learning-literature . Overall, this paper contributes significantly to understanding membership inference attacks on machine learning models by providing a comprehensive survey of existing research efforts and suggesting future research directions.