This repository contains implementation of Membership Inference, Attribute Inference and Model Inversion attacks as a part of course project for 'Privacy Enhancing Technologies' offered at Saarland University by Prof. Yang Zhang @ CISPA, Saarbrücken, Germany.
Course homepage | Project guidelines | Project references | Results and Inferences
Note: There are two implementation versions of the CLI. The other implementation can be found in 'CLI_ver2' folder.
$ python cli.py membership-inference [OPTIONS] COMMAND [ARGS]...
-
Commands:
pretrained-dummyLoad trained target and shadow model and train attack modeltrain-dummyTrain target, shadow and attack modeltrain-plus-dummyTrain target, shadow and attack model + augmentation, topk posteriors, parameter initialization and verbose enabled (check note below)
-
Options:
--dataset TEXTWhich dataset to use (CIFAR10 or MNIST) [default = CIFAR10]--data_path TEXTPath to store data [default = 'Membership-Inference/data']--model_path TEXTPath to save or load model checkpoints [default = 'Membership-Inference/model']--helpShow this message and exit.
-
Example commands:
-
$ python cli.py membership-inference pretrained-dummy --dataset MNIST --model_path Membership-Inference/best_models/ -
$ python cli.py membership-inference train-dummy -
$ python cli.py membership-inference train-plus-dummy --need_topk --param_init -
Note:
Membership-Inference/modelandMembership-Inference/datafolders are generated while training.train-plus-dummyhas additional optional options which works as flags if mentioned:--need_augmTo use data augmentation on target and shadow training set or not--need_topkFlag to enable using Top 3 posteriors for attack data--param_initFlag to enable custom model params initialization--verboseAdd Verbosity
-
$ python cli.py attribute-inference [OPTIONS] COMMAND [ARGS]...
-
Options:
--helpShow this message and exit.
-
Commands:
pretrained-dummyLoad trained target and attack modelsupply-targetSupply own target model and train attack modeltrain-dummyTrain target and attack model
-
$ python cli.py attribute-inference pretrained-dummy [OPTIONS]- Load trained target and attack model
-
Options:
--helpShow this message and exit.
-
Example command:
$ python cli.py attribute-inference pretrained-dummy
-
$ python cli.py attribute-inference train-dummy [OPTIONS]- Train target and attack model
-
Options:
-t,--target_epochsINTEGERNumber of training epochs for the target model [default=30]-a,--attack_epochsINTEGERNumber of training epochs for the attack model [default=50]--helpShow this message and exit.
-
Example command:
$ python cli.py attribute-inference train-dummy -t 30 -a 50
-
$ python cli.py attribute-inference supply-target [OPTIONS]- Supply own target model and train attack model.
- Specifications for the target Model:
- The target model should predict gender of a human [0: male, 1:female] and trained in UTKFace dataset.
- The target model should give the following output:
y, xwhere y are the two posteriors and y is the last fully connected layer. E.g.[ 5.0912e-01, -5.4544e-01], [-0.0656, 0.0087, -0.0543, ..., 0.0134, 0.0608, -0.0347] - The class file needs to be in the attribute inference folder.
- Name of the class needs to be 'TargetModel'.
-
Options:
-c,--class_fileTEXTFile that holds the target models nn.Module class [required]-s,--state_pathTEXTPath of the state dictionary [required]-d,--dimensionINTEGERFlattend dimension of the layer used as attack modelinput [required]-a,--attack_epochsINTEGERNumber of training epochs for the attack model [default=30]--helpShow this message and exit.
-
Example command:
$ python cli.py attribute-inference supply-target -c af_models -s Attribute-Inference/models/target_model_30.pth -d 64 -a 50
$ python cli.py model-inversion [OPTIONS] COMMAND [ARGS]...
-
Options:
--helpShow this message and exit.
-
Commands:
pretrained-dummyLoad trained target model and perform inversiontrain-dummyTrain target model and perform model inversionsupply-targetUse trained external target model and perform model inversion
-
$ python cli.py model-inversion pretrained-dummy [OPTIONS]- Load trained target model and perform inversion
- Pretrained model 'atnt-mlp-model.pth' is in the ModelInversion folder.
-
Options:
--iterations INTEGERNumber of Iterations in attack [default = 30]--loss_function [crossEntropy|softmax]which loss function to used crossEntropy or softmax [default = crossEntropy]--generate_specific_class INTEGERchoose class, number between 1 and 40, which you want recovered or nothing to get all recovered [default = -1]--helpShow this message and exit.
-
Example command:
$ python cli.py model-inversion pretrained-dummy
-
$ python cli.py model-inversion train-dummy [OPTIONS]- Train target model and perform model inversion
-
Options:
--iterations INTEGERNumber of Iterations in attack [default = 30]--epochs INTEGERNumber of epochs for the target model [default = 30]--loss_function [crossEntropy|softmax]which loss function to used crossEntropy or softmax [default = crossEntropy]--generate_specific_class INTEGERchoose class, number between 1 and 40, which you want recovered or nothing to get all recovered [default = -1]--helpShow this message and exit.
-
Example command:
$ python cli.py model-inversion train-dummy --epochs 30
-
$ python cli.py model-inversion supply-target [OPTIONS]- Use trained external target model and perform model inversion
- Specifications for the target Model:
- The target model should based on ATnT faces dataset.
- The target model needs to return 'output, h' where output are the posteriors (h can be neglected).
- The class file needs to be in the model inversion folder.
- Name of the class needs to be 'TargetModel'.
-
Options:
--class_file TEXTFile that holds the target models nn.Module class [required]--target_model_path TEXTtarget model file [required]--iterations INTEGERNumber of Iterations in attack [default = 30]--loss_function [crossEntropy|softmax]which loss function to used crossEntropy or softmax [default = crossEntropy]--generate_specific_class INTEGERchoose class, number between 1 and 40, which you want recovered or nothing to get all recovered [default = -1]--helpShow this message and exit.
-
Example command:
$ python cli.py model-inversion supply-target --class_file target_model --target_model_path ModelInversion/atnt-mlp-model.pth --generate_specific_class 25
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent