This repository includes the code to reproduce our experiments in the submission entitled
FLea: Addressing Data Scarcity and Label Skew in Federated Learning via Privacy-Preserving Feature Augmentation
accepted by KDD 2024 Research Track
To install requirements:
conda env create -f environment.yaml
conda activate FLea
conda env list
We mainly use torch, torchvision, torchaudio, numpy, sklearn,pandas
├── FLea
│ ├── data # data path
│ │ ├── CIFAR10
│ │ ├── UrbanSound8K
│ │ ├── UCI_HAR
│ ├── Utils # function path
│ │ ├── data_utils # data partition
│ │ ├── model_utils
│ │ ├── argparse
│ │ ├── functional
│ ├── FLAlgorithms # main algorithm path
│ │ ├── servers # code for server function
│ │ ├── users # code for client function
│ │ ├── trainmodel # model architecture
│ ├── run # script for running code
│ │ ├── Image
│ │ ├── Audio
│ │ ├── HAR
│ ├── models # saved models and logs
│ │ ├── saved
│ │ ├── logs
We use three public datasets, put in data path.
-
CIFAR10 will be downloaded automatically for the first run.
-
UrbanSound8K can be downloaded from UrbanSound8K - Urban Sound Datasets (weebly.com)
-
UCI_HAR can be downloaded from Human Activity Recognition Using Smartphones - UCI Machine Learning Repository
Official training and testing split are used in our experiments.
We split the training set data into K clients. To simulate label skew, we consider quantity-based skew (Qua(q) with q being the number of presented classes) and distribution-based skew (Dir(u) with u controlling the class skewness). The partition method is adapted from FedLab-NLP/fedlab/utils/dataset/functional.py. To mimic different levels of data scarcity, we distribute the training data to K clients, where K is determined by the average local data size, which is set to as small as 100 and 50 for experiments.
An example of the CIFAR-10 data distribution is visualized below. The training data for 10 classes are distributed across 100 clients, showcasing a quantity-based skew, with each client having only 3 classes of data. Each column represents one client, and the number of samples is denoted by color.
There are a few setups in our code:
--dataset: The dataset to use: 'Cifar10', 'UrbanSound', 'UCI_HAR' \
--split_method: The data particiption method: 'quantity','distribution' \
--split_para: Data particiption paraemter: 3, 0.5, 0.1 \
--split_num 1000: The number of clients: K \
--client_num 100: The number of clients per round \
--class_num: The number of classes \
--local_epochs: The number of epoches per round\
--batch_size: Batch size for local training: 32, 64 \
--num_global_iters: The total number of rounds: 100 \
--personal_learning_rate: Local learning rate： 0.001 \
--modelname: Model: 'MOBNET', 'AudioNet','HARNet' \
--algorithm: Our proposed FL method and baselines: 'FLea', 'FedAvg', 'FedProx', 'FedDecor','FedLC', 'FedNTD', 'FedBR', 'CCVR', 'FedMix_pre', 'FedData_pre'\
--loss: Local training lossfuction: 'MCE_DeC_KL', 'CE', 'MCE' \
--layer: From which CNN block the activations are eacted for FLea: 1 \
--fea_percent: The precentage of the data/featuers shared out: 0.1 \
--seed: random seed: 0, 1, 10, 100, 100, 1000 \
All the running scripts are provided in folder ./run/
To run the code for FLea, using CIFAR10 data, Dir(0.1), K=1000 (average local data size 50)
sh run/Image/run_FLea.sh
To run the code for FLea, using audio data, Qua(3), K= 70 (average local data size 100)
sh run/Audio/run_FLea_audio.sh
To run the code for FLea, using HAR data, Dir(0.1), K=150 (average local data size 50)
sh run/HAR/FLea_HAR.sh
Baseline implementations are included in our code base. We provide the running script for FedAvg, FedProx, FedDecor,FedLC, FedNTD, FedBR, CCVR, FedMix, FedData
To run the code for FedAvg, using CIFAR10 data, distribution skew, K=1000 (average local data size 50)
sh run/Image/run_FedAvg.sh
To run the code for FedMix,
sh run/Image/run_FedMix.sh
Our training logs can also be found from ./models/logs/
For example, the model trained for HAR data, using K=150, partition Dir(0.1) can be found from FLea_HAR_K150_Dir0.1.log
================================================================================
Summary of training process:
Dataset : UCI_HAR
Batch size : 32
Learing rate : 0.001
Number of total clients: 150
Split method : distribution
Split parameter : 0.1
Clients per round : 15
Number of global rounds: 100
Number of local rounds : 5
Feature from layer : 1
Feature percentage : 0.1
Is interplolated : False
Local training loss : MCE_DeC_KL
Loss of beta : 1.0
Algorithm : FLea
Modelname : HARNet
Mode : training
Seed : 0
================================================================================
Dataset: UCI_HAR
training: 7352 testing: 2947
.....
Number of users per round / total users: 15 / 150
Finished creating FL server.
=== Training starts: algorithm FLea ===
-------------Round number: 0 -------------
......
-------------Round number: 57 -------------
Global Model Acc on global data: 0.6599932134373939
save a model
......
Trained model checkpoints are saved in ./models/saved/
For example, the model trained for HAR data, using K=150, partition Dir(0.1) can be found from ./models/saved/server_FLea_HARNet_UCI_HAR_loss_MCE_DeC_KL_epoch_5_100_client_150_split_distribution_0.1.pt
Training and testing the data reconstruction attack: python test_reconstraction.py
