[ESORICS 24] VFLIP: A Backdoor Defense for Vertical Federated Learning via Identification and Purification

In support of open science, we will release the full implementation of VFLIP including SOTA attack methods used in the experiments to an open-source.

Currently, the evaluation code and trained models for VFLIP are provided.

Installation

Install required packages

$ pip install -r requirements.txt

Download zip files of pre-trained models & backdoor patterns for cifar via below link

Download Pre-trained models and patterns

Move the loaded zip files into ./toy_data directory or use --load-dir to specify your files's directory.

Unzip the loaded zip files in the directory

Inference Example

# BadVFL passive attack, a single attacker, 4 party scenario 
# "./toy_data/CIFAR10_BadVFL_ALL_False_4.pth"
$ python main.py -d CIFAR10 --path-dataset ./data/CIFAR10 --k 4 --gpu 0 --party-num 4 --defense-type VFLIP --attack-type BadVFL
$ python main.py -d CIFAR10 --path-dataset ./data/CIFAR10 --k 4 --gpu 0 --party-num 4 --defense-type NONE --attack-type BadVFL

# VILLAIN passive attack, a single attacker, 4 party scenario 
# "./toy_data/CIFAR10_VILLAIN_ALL_False_4.pth"
$ python main.py -d CIFAR10 --path-dataset ./data/CIFAR10 --k 4 --gpu 0 --party-num 4 --defense-type VFLIP --attack-type VILLAIN
$ python main.py -d CIFAR10 --path-dataset ./data/CIFAR10 --k 4 --gpu 0 --party-num 4 --defense-type NONE --attack-type VILLAIN

# BadVFL active attack, a single attacker, 4 party scenario 
# "./toy_data/CIFAR10_BadVFL_ALL_True_4.pth"
$ python main.py -d CIFAR10 --path-dataset ./data/CIFAR10 --k 4 --gpu 0 --party-num 4 --defense-type VFLIP --attack-type BadVFL --active
$ python main.py -d CIFAR10 --path-dataset ./data/CIFAR10 --k 4 --gpu 0 --party-num 4 --defense-type NONE --attack-type BadVFL --active

# VILLAIN active attack, a single attacker, 4 party scenario 
#"./toy_data/CIFAR10_VILLAIN_ALL_True_4.pth"
$ python main.py -d CIFAR10 --path-dataset ./data/CIFAR10 --k 4 --gpu 0 --party-num 4 --defense-type VFLIP --attack-type VILLAIN --active
$ python main.py -d CIFAR10 --path-dataset ./data/CIFAR10 --k 4 --gpu 0 --party-num 4 --defense-type NONE --attack-type VILLAIN --active

# BadVFL active attack with DP-SGD, a single attacker, 4 party scenario 
# "./toy_data/CIFAR10_BadVFL_DPSGD_True_4.pth"
$ python main.py -d CIFAR10 --path-dataset ./data/CIFAR10 --k 4 --gpu 0 --party-num 4 --defense-type DPSGD --attack-type BadVFL --active

# VILLAIN active attack with DP-SGD, a single attacker, 4 party scenario 
# "./toy_data/CIFAR10_VILLAIN_DPSGD_True_4.pth"
$ python main.py -d CIFAR10 --path-dataset ./data/CIFAR10 --k 4 --gpu 0 --party-num 4 --defense-type DPSGD --attack-type VILLAIN --active

# BadVFL active attack, 3 attackers, 8 party scenario 
# "./toy_data/CIFAR10_BadVFL_ALL_True_8.pth"
$ python main.py -d CIFAR10 --path-dataset ./data/CIFAR10 --k 4 --gpu 0 --party-num 8 --defense-type VFLIP --attack-type BadVFL --active --bkd-adversary 3,4,6
$ python main.py -d CIFAR10 --path-dataset ./data/CIFAR10 --k 4 --gpu 0 --party-num 8 --defense-type NONE --attack-type BadVFL --active --bkd-adversary 3,4,6

# VILLAIN active attack, 3 attackers, 8 party scenario 
# "./toy_data/CIFAR10_VILLAIN_ALL_True_8.pth"
$ python main.py -d CIFAR10 --path-dataset ./data/CIFAR10 --k 4 --gpu 0 --party-num 8 --defense-type VFLIP --attack-type VILLAIN --active --bkd-adversary 3,4,6
$ python main.py -d CIFAR10 --path-dataset ./data/CIFAR10 --k 4 --gpu 0 --party-num 8 --defense-type NONE --attack-type VILLAIN --active --bkd-adversary 3,4,6