Skip to content

Latest commit

 

History

History
97 lines (76 loc) · 6.05 KB

README.md

File metadata and controls

97 lines (76 loc) · 6.05 KB
Raw

Locally Private Graph Neural Networks

This repository is the official implementation of the paper:
Locally Private Graph Neural Networks (ACM CCS '21)

Proceedings version: https://dl.acm.org/doi/abs/10.1145/3460120.3484565
Video presentation: https://www.youtube.com/watch?v=1LdC5G_p-0g

Abstract

Graph Neural Networks (GNNs) have demonstrated superior performance in learning node representations for various graph inference tasks. However, learning over graph data can raise privacy concerns when nodes represent people or human-related variables that involve sensitive or personal information. In this paper, we study the problem of node data privacy, where graph nodes (e.g., social network users) have potentially sensitive data that is kept private, but they could be beneficial for a central server for training a GNN over the graph. To address this problem, we propose a privacy-preserving, architecture-agnostic GNN learning framework with formal privacy guarantees based on Local Differential Privacy (LDP). Specifically, we develop a locally private mechanism to perturb and compress node features, which the server can efficiently collect to approximate the GNN's neighborhood aggregation step. Furthermore, to improve the accuracy of the estimation, we prepend to the GNN a denoising layer, called KProp, which is based on the multi-hop aggregation of node features. Finally, we propose a robust algorithm for learning with privatized noisy labels, where we again benefit from KProp's denoising capability to increase the accuracy of label inference for node classification. Extensive experiments conducted over real-world datasets demonstrate that our method can maintain a satisfying level of accuracy with low privacy loss.

figure

Requirements

This code is implemented in Python 3.9, and relies on the following packages:

Note: For the DGL-based implementation, switch to the DGL branch.

Usage

Replicating the paper's results

In order to replicate our experiments and reproduce the paper's results, you must do the following steps:

  1. Run python experiments.py -n LPGNN create --LPGNN --baselines
  2. Run python experiments.py -n LPGNN exec --all
    All the datasets will be downloaded automatically into datasets folder, and the results will be stored in results directory.
  3. Go through results.ipynb notebook to visualize the results.

Training individual models

If you want to individually train and evaluate the models on any of the datasets mentioned in the paper, run the following command:

python main.py [OPTIONS...]

dataset arguments:
  -d              <string>       name of the dataset (choices: cora, pubmed, facebook, lastfm) (default: cora)
  --data-dir      <path>         directory to store the dataset (default: ./datasets)
  --data-range    <float pair>   min and max feature value (default: (0, 1))
  --val-ratio     <float>        fraction of nodes used for validation (default: 0.25)
  --test-ratio    <float>        fraction of nodes used for test (default: 0.25)

data transformation arguments:
  -f              <string>       feature transformation method (choices: raw, rnd, one, ohd) (default: raw)
  -m              <string>       feature perturbation mechanism (choices: mbm, 1bm, lpm, agm) (default: mbm)
  -ex             <float>        privacy budget for feature perturbation (default: inf)
  -ey             <float>        privacy budget for label perturbation (default: inf)

model arguments:
  --model         <string>       backbone GNN model (choices: gcn, sage, gat) (default: sage)
  --hidden-dim    <integer>      dimension of the hidden layers (default: 16)
  --dropout       <float>        dropout rate (between zero and one) (default: 0.0)
  -kx             <integer>      KProp step parameter for features (default: 0)
  -ky             <integer>      KProp step parameter for labels (default: 0)
  --forward       <boolean>      applies forward loss correction (default: True)

trainer arguments:
  --optimizer     <string>       optimization algorithm (choices: sgd, adam) (default: adam)
  --max-epochs    <integer>      maximum number of training epochs (default: 500)
  --learning-rate <float>        learning rate (default: 0.01)
  --weight-decay  <float>        weight decay (L2 penalty) (default: 0.0)
  --patience      <integer>      early-stopping patience window size (default: 0)
  --device        <string>       desired device for training (choices: cuda, cpu) (default: cuda)

experiment arguments:
  -s              <integer>      initial random seed (default: None)
  -r              <integer>      number of times the experiment is repeated (default: 1)
  -o              <path>         directory to store the results (default: ./output)
  --log           <boolean>      enable wandb logging (default: False)
  --log-mode      <string>       wandb logging mode (choices: individual,collective) (default: individual)
  --project-name  <string>       wandb project name (default: LPGNN)

The test result for each run will be saved as a csv file in the directory specified by
-o option (default: ./output).

Citation

If you find this code useful, please cite the following paper:

@inproceedings{sajadmanesh2021locally,
   author = {Sajadmanesh, Sina and Gatica-Perez, Daniel},
   title = {Locally Private Graph Neural Networks},
   year = {2021},
   publisher = {Association for Computing Machinery},
   doi = {10.1145/3460120.3484565},
   booktitle = {Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security},
   pages = {2130–2145},
   series = {CCS '21}
}