Code for MixTailor: Mixed Gradient Aggregation for Robust Learning Against Tailored Attacks to appear in TMLR.
We recommend using Anaconda to install the following packages:
- Python 3.7.1
- PyTorch
- TensorboardX
- Pyyaml
- Scipy
conda install pytorch torchvision -c pytorch
pip install pyyaml scipy tensorboardx
The code mainly cosists of the implementation of aggregators and attacks. When the number of byzantine workers is larger than 0, the attack API allows to return any custom gradient based on the gradient of non-byzantine workers. In this framework, we first calculate the gradient of honest workers based on the current data sample, then we provide these gradients to the byzantine work and retrieve the attack gradient, finally we pass all the honest and byzantine gradients to the aggregator. The aggregator produces the final aggregated gradient which will be applied to the model.
All the attacks are implemented in dbqf/attacks.py.
To implement your own custom attacks you need to extend the Attack class and implement the grad function. The grad function
receives a list of honest gradients and has to return the attack gradient.
The aggregator receives a pool of graidents where some of them come from honest workers and others come from byzantine workers.
The main job of the aggregator is to find out which aggregators are honest and which ones are not and produce the final gradient based on that.
The gradient produced by the aggregator will be applied to the model. All the aggregators are implemented in dbqf/aggregators.py. In order to implement your own aggregator, you need to
extend the Aggregator class and implement the agg_g function. This function receives a list of gradients (some honest, some byzantine) and
must return a tensor with the same size that contains the aggregated final gradient.
@article{ramezani2022mixtailor,
title={MixTailor: Mixed Gradient Aggregation for Robust Learning Against Tailored Attacks},
author={Ramezani-Kebrya, Ali and Tabrizian, Iman and Faghri, Fartash and Popovski, Petar},
journal={arXiv preprint arXiv:2207.07941},
year={2022}
}