SoK: Comparing Different Membership Inference Attacks with a Comprehensive Benchmark

Jun Niu, Xiaoyan Zhu, Moxuan Zeng, Ge Zhang, Qingyang Zhao, Chunhui Huang, Yangming Zhang, Suyu An, Yangzhong Wang, Xinghui Yue, Zhipeng He, Weihao Guo, Kuo Shen, Peng Liu, Yulong Shen, Xiaohong Jiang, Jianfeng Ma, Yuqing ZhangarXiv 2023 · DOI · arXiv

Introduction

Membership Inference (MI) attacks represent a significant privacy risk in machine learning, allowing an adversary to determine if a specific data record was used to train a target model. This is particularly concerning when models are trained on sensitive data, such as medical records or private financial information. Despite the proliferation of research in this area, comparing the effectiveness of different attacks has remained a challenge due to inconsistent evaluation environments and a lack of standardized benchmarks.

The "Comparison Crisis" in MI Research

The authors of this study argue that MI research is currently in a "comparison crisis." Most existing literature compares new attacks against older ones using disparate datasets, varying model architectures, and inconsistent evaluation metrics. Crucially, these comparisons often ignore the geometry of the data. The researchers found that the perceived strength of an attack is frequently more dependent on the internal distances within a dataset than on the algorithm itself. Without accounting for these factors, it is impossible to determine which attacks are truly the most effective.

MIBench: A Standardized Framework

To resolve these inconsistencies, the authors introduced MIBench, a comprehensive benchmark designed to systematize the evaluation of MI attacks. MIBench evaluates 15 state-of-the-art (SOTA) attacks across 588 distinct scenarios.

The attack workflow within MIBench, illustrating how target datasets are processed through various MI attack algorithms to produce membership inferences.

The core of MIBench's innovation lies in its use of four Control Variables (CVs) to create standardized evaluation scenarios:

  1. CV1 (Distance Distribution): Standardizing how data samples are distributed (e.g., Normal, Uniform, or Bernoulli).
  2. CV2 (Distance Between Samples): Using Maximum Mean Discrepancy (MMD) to measure the "spread" of the target dataset.
  3. CV3 (Differential Distance): Measuring the change in MMD when a sample is moved between the target and non-member sets.
  4. CV4 (Ratio of No Inferences): Accounting for "fine-grained" attacks that maintain high precision by choosing not to make predictions on uncertain samples.

Key Insights: Geometry is Destiny

A primary finding of the study is that the effectiveness of an MI attack is highly sensitive to the distance between data samples. As the distance between samples increases, the "Membership Advantage" (the gap between true and false positives) grows significantly across all tested attack types.

Experimental results on CIFAR10 showing a clear positive correlation between the distance between data samples and the Membership Advantage across multiple attack algorithms.

This geometric dependency leads to what the authors call the "Ranking Flip": an attack that appears superior on a dataset with a specific distribution may perform worse than its competitors when the distribution changes. This explains why different research papers often report contradictory results. Furthermore, the study revealed that "differential distance" (CV3) generally exerts a greater influence on attack effectiveness than simple sample distance (CV2).

Results and Implications

The benchmarking process quantified the impact of these variables. For instance, on the CIFAR100 dataset, the Membership Advantage of a standard neural network-based attack jumped from 28.89% to 39.39% simply by increasing the distance between data samples, without any modification to the attack algorithm itself.

By providing 10 typical evaluation metrics (including Accuracy, Precision, Recall, and AUC), MIBench ensures a multi-dimensional view of privacy risk. The authors' review of 15 major MI papers found that none met all the requirements for a fair comparison, and over half failed to meet even one, highlighting the urgent need for the standardization provided by MIBench.

Conclusion

MIBench levels the playing field for Membership Inference research by isolating the algorithmic performance from the underlying data geometry. Future work will look to integrate standardized defense mechanisms, such as Differential Privacy, into the framework and extend these geometric control variables to more complex data structures like graphs and time-series data.

Summary updated

Abstract

Membership inference (MI) attacks threaten user privacy through determining if a given data example has been used to train a target model. However, it has been increasingly recognized that the "comparing different MI attacks" methodology used in the existing works has serious limitations. Due to these limitations, we found (through the experiments in this work) that some comparison results reported in the literature are quite misleading. In this paper, we seek to develop a comprehensive benchmark for comparing different MI attacks, called MIBench, which consists not only the evaluation metrics, but also the evaluation scenarios. And we design the evaluation scenarios from four perspectives: the distance distribution of data samples in the target dataset, the distance between data samples of the target dataset, the differential distance between two datasets (i.e., the target dataset and a generated dataset with only nonmembers), and the ratio of the samples that are made no inferences by an MI attack. The evaluation metrics consist of ten typical evaluation metrics. We have identified three principles for the proposed "comparing different MI attacks" methodology, and we have designed and implemented the MIBench benchmark with 84 evaluation scenarios for each dataset. In total, we have used our benchmark to fairly and systematically compare 15 state-of-the-art MI attack algorithms across 588 evaluation scenarios, and these evaluation scenarios cover 7 widely used datasets and 7 representative types of models. All codes and evaluations of MIBench are publicly available at https://github.com/MIBench/MIBench.github.io/blob/main/README.md.
BibTeX
@misc{niu2023sok,
  author = {Jun Niu and Xiaoyan Zhu and Moxuan Zeng and Ge Zhang and Qingyang Zhao and Chunhui Huang and Yangming Zhang and Suyu An and Yangzhong Wang and Xinghui Yue and Zhipeng He and Weihao Guo and Kuo Shen and Peng Liu and Yulong Shen and Xiaohong Jiang and Jianfeng Ma and Yuqing Zhang},
  title = {{SoK: Comparing Different Membership Inference Attacks with a Comprehensive Benchmark}},
  year = {2023},
  eprint = {2307.06123},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2307.06123},
}
Figure