Skip to content

Latest commit

 

History

History
66 lines (47 loc) · 3.34 KB

README.md

File metadata and controls

66 lines (47 loc) · 3.34 KB
Raw

Label Leakage from Gradients

This framework implements the Label Leakage from Gradients (LLG) attack, a novel attack to extract ground-truth labels from shared gradients trained with mini-batch stochastic gradient descent for multi-class classification in Federated Learning. LLG is based on a combination of mathematical proofs and heuristics derived empirically. The attack exploits two properties that the gradients of the last layer of a neural network have: (P1) The direction of these gradients indicates whether a label is part of the training batch. (P2) The gradient magnitude can hint towards the number of occurrences of a label in the batch.

References

[1] Aidmar Wainakh et al. (2022) User-Level Label Leakage from Gradients in Federated Learning. In PETS 2022.

[2] Aidmar Wainakh et al. (2021) Label Leakage from Gradients in Distributed Machine Learning. In CCNC 2021.

Installation

  1. Setup a clean Python 3.7.9 environment with the tool of your choice (conda, venv, etc.).
  2. Install required python libraries using: pip install -r Code/requirements.txt
  3. Initiate and update aDPtorch submodule: git submodule init and git submodule update

It is possible that the LLG code runs with newer python versions. However, don't use the most current, as opacus and torchcsprng tend to have a bit of a delay getting updated to work with newest python and/or torch versions.

Execution

  1. Choose an experiment from the table below.
  2. Prepare the detailed experiment parameters in main.py to fit your needs.
  3. Execute the experiment: python main.py -s <experiment_set_number> -g <gpu_id_if_avail>
  4. Visualize the dump file(s): python main.py -s <experiment_set_number> -d <path_to_dump_file(s)>

Experiment Sets

set description
1,2 batch size (untrained)
3,4 trained model
5 model architecture comparison
6 additive noise (untrained)
7 compression (untrained)
8 differential privacy (untrained)
9 federated training and trained defenses

Current CLI

usage: main.py [-h] [-s SET] [-p PLOT] [-j JOB] [-d DIR] [-g GPU_ID]

Arguments for LLG Experiment

optional arguments:
  -h, --help                    show this help message and exit
  -s SET, --set SET             experiment set (default=2)
  -p PLOT, --plot PLOT          number of files to be ploted (default=None)
  -j JOB, --job JOB             job to execute. either "experiment" or "visualize". (default="experiment")
  -d DIR, --dir DIR             directory or file to plot from. (default=None)
  -g GPU_ID, --gpu_id GPU_ID    cuda_id to use, if available (default=0)

Contributors

  • Aidmar Wainakh - LLG idea, guidance and suggestions during development
  • Till Müßig - LLG idea, developing LLG and initial experiments as part of his Bachelor’s thesis and a seminar course
  • Jens Keim - developing advanced experiments, refactoring, current maintainer

License

This repository is licensed under the MIT License.

This repo contains a markdown and a text version of the license.

In case of any inconstancies refer to the license's website.