Project introduction

This is the official repository for the ACM CCS 2024 paper "SeqMIA: Sequential-Metric Based Membership Inference Attack" by Hao Li, Zheng Li, Siyuan Wu, Chengrui Hu, Yutong Ye, Min Zhang, Dengguo Feng, and Yang Zhang. In this project, we demonstrate the Sequential-metric based Membership Inference Attack(SeqMIA) that recognizes and utilizes integrated membership signal: the Pattern of Metric Sequence, derived from the various stages of model training.

Environment dependencies

pytorch:2.0.1+cu118

python:3.8.17

Note that we tested this under Windows. If you are on Linux, you may need to change the matplotlib plotting part.

Directory structure

  ├── README.md
  
  ├── data
  
  │   ├── cifar-10-batches-py-official
  
  │   ├── CIFAR10
  
  │       ├── PreprocessedIncludingDistillationData
  
  ├── model_IncludingDistillation
  
  │   ├── CIFAR10
  
  │       ├── Shadow
  
  │       ├── Target
  
  ├── results
  
  ├── attackMethodsFramework.py
  
  ├── Metrics.py
  
  ├── MetricSequence.py
  
  ├── Models.py
  
  ├── readData.py
  
  ├── SeqMIA.py
  
  └── temp

Supported Dataset and Model

CIFAR10

vgg-16

Following this example, more datasets and models can be added to SeqMIA.

Usage instructions

First, please put the CIFAR10 dataset (downloaded from its official web, i.e. data_batch_1 to data_batch_5 and test_batch) into cifar-10-batches-py-official. Then, you can run attackMethodsFramework.py to start the entire attack process.

You can set some hyperparameters in set_args() or change them on the command line as follows.

1.The first time you run this code, please set preprocessData to True. And processed data will be in the PreprocessedIncludingDistillationData folder.

2.Set trainTargetModel and trainShadowModel to True to train the target model and the shadow model when needed.

3.Set distillTargetModel and distillShadowModel to True to distill the models when needed.

For example, after placing the CIFAR10 dataset into cifar-10-batches-py-official, run the command python attackMethodsFramework.py --preprocessData True --trainTargetModel True --trainShadowModel True --distillTargetModel True --distillShadowModel True.

Authors

This project is created by Hao Li (ISCAS)

Cite

If you find this work useful in your research, please consider citing SeqMIA: Sequential-Metric Based Membership Inference Attack

@inproceedings{li2024seqmia,
    author = {Hao Li and Zheng Li and Siyuan Wu and Chengrui Hu and Yutong Ye and Min Zhang and Dengguo Feng and Yang Zhang},
    title = {{SeqMIA: Sequential-Metric Based Membership Inference Attack}},
    booktitle = {{ACM SIGSAC Conference on Computer and Communications Security (CCS)}},
    publisher = {ACM},
    year = {2024}
}