human-analysis / closed-form-adversarial-representation-learning

Adversarial Representation Learning With Closed-Form Solvers (ECML 2021), Pytorch Implementation

Adversarial Representation Learning with Closed-Form Solvers

Requirements

1. Require Python3
2. Require PyTorch1.0
3. Require Visdom0.1.8.9
4. Check requirements.txt for detailed dependencies.

Commands to Reproduce Results in Paper

Synthetic Gaussian Dataset

SGDA-ARL

$ python3 -m visdom.server
$ python3 main_gaussian.py --args args/Gaussian-SGDA-ARL.txt

OptNet-ARL

$ python3 -m visdom.server
$ python3 main_gaussian.py --args args/Gaussian-OptNet-ARL.txt

CelebA Dataset

SGDA-ARL

$ python3 -m visdom.server
$ python3 main_celebA.py --args args/CelebA-SGDA-ARL.txt

OptNet-ARL

$ python3 -m visdom.server
$ python3 main_celebA.py --args args/CelebA-OptNet-ARL.txt

Part A: Training the Encoder

1. Set the path to your input data and your dataset name for both training and test sets. Note: Let the data created by dataloader.py contain three items, input data, target class label and sensitive class label, respectively. Example in args/CelebA-OptNet-ARL.txt:

   dataset_root_test = ./data/celeba/
   dataset_root_train = ./data/celeba/
   dataroot = ./data/celeba/
   
   dataset_train = CelebA_Privacy
   dataset_test = CelebA_Privacy
   
   input_filename_train = ./data/celeba/celeba-training.csv
   input_filename_test = ./data/celeba/celeba-evaluation.csv
   
   

2. Set the dimentionality of your embedding r and datandim, number of sensitive class label nclasses_A, and number of target class label nclasses_T. Example in args.txt:

   r = 2
   #### due to instant normalization one dimension will be lost
   
   resolution_high = 112
   resolution_wide = 96
   nclasses_A = 100
   nclasses_T = 20
   

3. Set a set trade-off parameters (0<=alpha<=1) between privacy and utility. Note: alpha=[0] is related to no privacy and alpha=[1] concerns totally to hide the sensitive attribute.

   alpha = [0, 0.1, 0.3 ,0.5, 0.7, 0.8, 0.85, 0.9, 0.91, 0.92, 0.93, 0.94]
   
   

4. Choose your ARL method and associated networks.

   Example for OptNet-ARL:

   adverserial_type = OptNet
   loss_type_E = Projection_gauss
   sigma = 1
   
   

   Example for SGDA-ARL:

   adverserial_type = SGDA
   model_type_EA = EA
   model_type_ET = ET
   

Part B: Training the Real Adversary and Target Classifiers or Regressors

1. Visualization Settings. The parameters for visdom to plot training and testing curves.

    1) the port number for visdom -- "port"
    2) the name for current environment -- "env"
    3) if you want to create a new environment every time you run the program or
     not -- "same_env".  If you do, set it "False"; otherwise, it's "True".
   

   Example in args.txt:

   port = 8097
   env = main
   same_env = True
   

2. Select the network for target and adversary and specify their task as a regression or classification. Example in args.txt:

   model_type_A = Adversary
   model_type_T = Target
   loss_type_A = Regression
   loss_type_T = Regression
   evaluation_type_A = Top1Classification
   evaluation_type_T = Top1Classification
   

3. Finally, set the hyper parameters required to train and test the real adversary and target networks. Example in args.txt:

   nepochs = 7
   optim_method = Adam
   learning_rate_T = 3e-4
   learning_rate_A = 3e-4
   scheduler_method_A = ExponentialLR
   scheduler_options_A = {"gamma": 0.999}