The poisoning-attacks-wireless-traffic-prediction from iqua

poisoning-attacks-wireless-traffic-prediction's Introduction

Acknowledgment

The code is written based on the Zhang et al's code, and still under polishment.

Download the dataset

The dataset can be manually downloaded from Google Drive

After downloading the .h5 files, put those files into a directory named dataset and put the directory under the main directory before running the code.

Train centralized model

python centralized.py --file milano.h5 --type sms --lr 1e-2 --frac 0.1 --bs 100 --opt 'sgd' --out_dim 1 --epochs 10 --batch_size 50

Centrailized File: milano.h5 Type: sms MSE: 0.3171 MAE: 0.3223, NRMSE: 0.0908

Uniform noise for comparison

python centralized.py --file milano.h5 --type sms --lr 1e-2 --frac 0.1 --bs 100 --opt 'sgd' --out_dim 1 --poison --attack_epsilon 0.2 --attack_optimizer uniform

Data poisoning against centralized model training

python centralized.py --file milano.h5 --type sms --lr 1e-2 --frac 0.1 --bs 100 --opt 'sgd' --out_dim 1 --poison --attack_epsilon 0.2 --num_ensemble 2 --attack_rounds 30 --epochs 10 --batch_size 50 --attack_lr 10.0 --mask_prob 0.8 --surrogate_model lstm

Apply defense to data poisoning (Data Sanitization, Randomized Smoothing)

python centralized.py --file milano.h5 --type sms --lr 1e-2 --frac 0.1 --bs 100 --opt 'sgd' --out_dim 1 --poison --attack_epsilon 0.2 --num_ensemble 2 --attack_rounds 30 --epochs 10 --batch_size 50 --attack_lr 10.0 --mask_prob 0.8 --surrogate_model lstm --apply_defense sphere_sani

python centralized.py --file milano.h5 --type sms --lr 1e-2 --frac 0.1 --bs 100 --opt 'sgd' --out_dim 1 --poison --attack_epsilon 1.0 --num_ensemble 2 --attack_rounds 20 --epochs 10 --batch_size 50 --attack_lr 10.0 --mask_prob 0.8 --surrogate_model lstm --apply_defense adj_sani

python centralized.py --file milano.h5 --type sms --lr 1e-2 --frac 0.1 --bs 100 --opt 'sgd' --out_dim 1 --poison --attack_epsilon 1.0 --num_ensemble 2 --attack_rounds 20 --epochs 10 --batch_size 50 --attack_lr 10.0 --mask_prob 0.8 --surrogate_model lstm --apply_defense rand

Train FedAvg model

python fed_avg.py --file milano.h5 --type sms --lr 1e-2 --frac 0.1 --bs 100 --opt 'sgd' --out_dim 1

Type: sms MSE: 0.3744 MAE: 0.3386, NRMSE: 0.0955

Model poisoning against FedAvg

python fed_avg.py --file milano.h5 --type sms --lr 1e-2 --frac 0.1 --bs 100 --opt 'sgd' --out_dim 1 --poison

Apply defenses (Multi Krum, Trimmed Mean)

python fed_avg.py --file milano.h5 --type sms --lr 1e-2 --frac 0.1 --bs 100 --opt 'sgd' --out_dim 1 --poison --apply_defense multi_krum

python fed_avg.py --file milano.h5 --type sms --lr 1e-2 --frac 0.1 --bs 100 --opt 'sgd' --out_dim 1 --poison --apply_defense trimmed_mean

python fed_avg.py --file milano.h5 --type sms --lr 1e-2 --frac 0.1 --bs 100 --opt 'sgd' --out_dim 1 --poison --apply_defense median

poisoning-attacks-wireless-traffic-prediction's People

Contributors

Stargazers

poisoning-attacks-wireless-traffic-prediction's Issues

Reproducibility issue

Hi @tianzheng4,
Thanks for sharing the code and congratulations for your interesting work.
However, it seems that we still could not control all the randomness in the code with seed settings.
On the conda environment created from the provided environment.yml, different runs might result in different MSE, MAE, NRMS values, especially when attacks and defenses are involved.
Do you have any suggestions here?
Thank you!

Package versions

Hi @tianzheng4 ,
Thanks for sharing this repository.
I'm having some compatibility issues with the package versions installed.
Could you please share the versions of packages or the .yml file of the environment required to run the scripts?
Many thanks!

No module named 'redundant.train_utils'

Hi, I'm getting this error. There's some modules missing:

from redundant.train_utils import normal_weight_init
ModuleNotFoundError: No module named 'redundant.train_utils'

Recommend Projects

iqua / poisoning-attacks-wireless-traffic-prediction Goto Github PK