This repository contains the source code of the IJCAI-22 paper "Membership Inference via Backdooring". The proposed approach is "MIB: Membership Inference via Backdooring".
- torch==1.8.1
- numpy==1.18.1
- torchvision==0.9.1
The experiments are evaluated on one image dataset of CIFAR-10 and two binary datasets of Location-30 and Purchase-100, which are widely used to evaluate membership privacy risks.
- For CIFAR-10, you can directly run the code and the dataset will be downloaded automatically.
- For Location-30, please first dowanlad it from here, and then put it in the "data" subfolder of the "Location" folder
- For Purchase-100, please first dowanlad it from here, and then put it in the "data" subfolder of the "Purchase" folder
python train_clean.py --gpu-id 0 --checkpoint 'checkpoint/benign_model'
- One data owner's data was collected and used: the default trigger pattern is a 3x3 white square and stamped in the bottom right of the selected samples. You can vary different --y_target, --trigger_size, and --marking_ratio to see how these factors affact the backdoor attack success rate. Note that adjusting the coordinate of the trigger.
python train_MIB.py --gpu-id 0 --checkpoint 'checkpoint/one_owner' --trigger 'white_square' --y_target 1 --trigger_size 3 --trigger_coordinate_x 29 --trigger_coordinate_y 29 --marking_rate 0.001
- Multiple data owners' data was collected and used: You can vary the number of data owners by changing --num_users. In the experiments, each data owner uses a different thrigger pattern and a different target label.
python train_MIB_multi.py --gpu-id 0 --checkpoint 'checkpoint/multi_owner' --num_users 10
python train_clean.py --gpu-id 0 --checkpoint 'checkpoint/benign_model'
- One data owner's data was collected and used: the default trigger is a 20-length binary array with each element of 1. The trigger pattern is placed in the end of the selelcted samples. You can vary different --y_target, --trigger_size, and --marking_ratio to see how these factors affact the backdoor attack success rate. Note that adjusting the coordinate of the trigger.
python train_MIB.py --gpu-id 0 --checkpoint 'checkpoint/one_owner' --trigger 'binary_1' --y_target 1 --trigger_size 20 --trigger_locate 426 --marking_rate 0.002
- Multiple data owners' data was collected and used: You can vary the number of data owners by changing --num_users. In the experiments, each data owner uses a different thrigger pattern and a different target label.
python train_MIB_multi.py --gpu-id 0 --checkpoint 'checkpoint/multi_owner' --num_users 10
python train_clean.py --gpu-id 0 --checkpoint 'checkpoint/benign_model'
- One data owners' data was collected and used: the default trigger is a 20-length binary array with each element of 1. The trigger pattern is placed in the end of the selelcted samples. You can vary different --y_target, --trigger_size, and --marking_ratio to see how these factors affact the backdoor attack success rate. Note that adjusting the coordinate of the trigger.
python train_MIB.py --gpu-id 0 --checkpoint 'checkpoint/one_owner' --trigger 'binary_1' --y_target 1 --trigger_size 20 --trigger_locate 580 --marking_rate 0.001
- Multiple data owner's data was collected and used: You can vary the number of data owners by changing --num_users. In the experiments, each data owner uses a different thrigger pattern and a different target label.
python train_MIB_multi.py --gpu-id 0 --checkpoint 'checkpoint/multi_owner' --num_users 10
Part of our code is based on the open-source code of the paper "Open-sourced Dataset Protection via Backdoor Watermarking", where backdooring technique was used to protect the intellectual property of datasets. We thank the contributions of the authors of that paper.