Task-Agnostic Privacy-Preserving Representation Learning for Federated Learning Against Attribute Inference Attacks

Implementation of the vanilla federated learning paper : Communication-Efficient Learning of Deep Networks from Decentralized Data. Updates done to implement a Privacy Preserving network.

Running the experiments

Federated experiment involves training a global model using many local models.

• Before running these experiments, it is necessary to install the following libraries:

 sudo pip3 install numpy
 sudo pip3 install pandas
 sudo pip3 install tqdm
 sudo pip3 install torch
 sudo pip3 install tensorboardX
 sudo pip3 install sklearn
 sudo pip3 install torchvision

EXPERIMENTS ON CIFAR-10

• To run the federated experiment with CIFAR on CNN (IID):

sudo python3 src/federated_main.py  --dataset=cifar --gpu=0 --iid=1 --num_classes=10

• To run the same experiment under non-IID condition:

sudo python3 src/federated_main.py  --dataset=cifar --gpu=0 --iid=0 --num_classes=10

EXPERIMENTS ON CREDIT For this dataset, two different private attributes can be defined: Gender (--private_attr=0) and Geography (--private_attr=1). For Geography, add the option --num_classes_priv=3 to the lines below.

• To run the federated experiment with credit dataset (IID):

sudo python3 src/federated_main.py  --dataset=credit --gpu=0 --iid=1

• To run the same experiment under non-IID condition:

sudo python3 src/federated_main.py  --dataset=credit --gpu=0 --iid=0

EXPERIMENTS ON LOANS For this dataset, two different private attributes can be defined: Gender (--private_attr=0) and Race (--private_attr=1).

• To run the federated experiment with credit dataset (IID):

sudo python3 src/federated_main.py  --dataset=loans --gpu=0 --iid=1 --num_classes=3

• To run the same experiment under non-IID condition:

sudo python3 src/federated_main.py  --dataset=loans --gpu=0 --iid=0 --num_classes=3

EXPERIMENTS ON ADULT INCOME For this dataset, two different private attributes can be defined: Gender (--private_attr=0) and Marital status (--private_attr=1). For Marital status, add the option --num_classes_priv=7 to the lines below.

• To run the federated experiment with credit dataset (IID):

sudo python3 src/federated_main.py  --dataset=adult_income --gpu=0 --iid=1

• To run the same experiment under non-IID condition:

sudo python3 src/federated_main.py  --dataset=adult_income --gpu=0 --iid=0