Quantile-based Bias Initialization for Efficient Private Data Reconstruction in Federated Learning

Requirements

We recommend setting up a dedicated environment with Python 3.11 then

pip install -r requirements.txt

Quick demo

For a quick demo including visualizations we provide two jupyter notebooks: Quick_Demo_Attack_(QBI).ipynb and Quick_Demo_Defense_(AGGP).ipynb.

Run them via jupyter notebook or jupyter lab -> open notebook -> Run all cells.

(~3s runtime for both, running on CPU only, CIFAR-10 dataset will be downloaded automatically)

ImageNet demo (Local dataset required)

For demos on the ImageNet dataset, we provide Demo_Attack_ImageNet_(QBI).ipynb and Demo_Defense_ImageNet_(AGGP).ipynb. (~15s runtime each)

Datasets

The CIFAR-10 and IMDB datasets will be downloaded automatically when an experiment is run, and saved in the datadirectory. To run experiments on ImageNet, the dataset has to be manually downloaded from image-net.org and placed in the data directory.

Running experiments

The files exp_{DATASET}_{METHOD}_<Normalization>.py contain the code to reproduce our results reported in Table 1-6, and can be run without any arguments. They will produce a file called results_{DATASET}_{METHOD}.csv in the results directory.

E.g. use the following command to run the experiments on CIFAR-10 with QBI:

python exp_CIFAR-10_QBI.py

Metrics

The following metrics (as defined in Section 4) are reported in the results files:

• A: Active neurons
• P: Extraction-Precision
• R: Extraction-Recall (Percentage of perfectly reconstructed samples)

95% confidence intervals across the 10 runs are also reported.

Customizing experiments

The default experiments will run each (layer_size,batch_size) combination with 10 different random initializations. Adjust the variables runs_per_setting, layer_sizes or batch_sizes, initialized in the main method of each experiment, to your liking.

Supplementary

Contained in the supplementary directory are the following files:

1. conv2d_identity.py Minimal working example for conv layers working as identity functions on multi-channel data, as described in Algorithm 3.
2. overlay2-32.py Script that produced Figure 5 (effect of averaging increasing number of samples)
3. similarity_metrics.py Script that produced Figure 6 (similarity metrics of increasing number of samples)