DeCart / DeCart* Experimental Repository

This repository includes implementations of DeCart, DeCart*, and multiple baseline schemes (CCS23/Server/Offline/SecPQ), along with experiment runners and paper plotting scripts.

1. Repository Overview

• schemes/: Core scheme implementations (DeCart, DeCart*, AI model abstraction)
• entities/: Participant entities (owner, querier, server, curator)
• core/: Finite field, bilinear pairing, homomorphic encryption, and other core modules
• experiments/: Runners for each scheme, dataset loaders, model training, and comparison plotting
• tests/: Basic functional and scheme comparison tests
• data/: MNIST/UCI HAR dataset directory

2. Environment Requirements

• Python 3.10 or 3.11 (recommended)
• Windows / Linux / macOS
• A virtual environment is recommended

3. Install Dependencies

Run in the project root:

python -m venv .venv
# Windows
.venv\Scripts\activate
# Linux/macOS
# source .venv/bin/activate

python -m pip install --upgrade pip
python -m pip install numpy sympy bn256 tenseal phe pycryptodome cryptography loguru matplotlib torch torchvision scikit-learn scipy

Notes:

• If you only run core schemes (no training or plotting), you may skip torch/torchvision/matplotlib/scikit-learn/scipy.
• tenseal may require a compatible build environment. Check Python and platform compatibility if installation fails.

4. Quick Start

4.1 Minimal Health Check

python tests/test_schemes_comparison.py

4.2 Run DeCart

python -m experiments.our_decart.runner --N 10000 --n 32 --num-records 32 --record-dim 32 --policy-size 32 --num-runs 3

4.3 Run DeCart*

python -m experiments.our_decart_star.runner --N 10000 --n 32 --num-records 32 --record-dim 32 --policy-size 32 --num-runs 3

Common arguments:

• --dataset {synthetic,mnist,uci_har}
• --model-source {synthetic,trained}
• --model-types dot decision_tree neural_network
• --no-save (do not persist results)
• --results-dir <path> (custom output directory)

5. Dataset and Dimension Constraints

• synthetic: record_dim is configurable
• mnist: record_dim must be 784
• uci_har: record_dim must be 561

It is recommended to start with the synthetic dataset.

python -m experiments.our_decart.runner --dataset synthetic --record-dim 32 --num-records 32 --policy-size 32 --num-runs 3
python -m experiments.our_decart_star.runner --dataset synthetic --record-dim 32 --num-records 32 --policy-size 32 --num-runs 3

To switch to real datasets:

python -m experiments.our_decart.runner --dataset mnist --record-dim 784 --num-records 128 --mnist-data-dir data
python -m experiments.our_decart_star.runner --dataset uci_har --record-dim 561 --num-records 128 --mnist-data-dir data

6. Train Real Models

python -m experiments.models.train_models --dataset mnist --data-dir data
python -m experiments.models.train_models --dataset uci_har --data-dir data

Trained models are saved to:

• experiments/models/trained/

Use them with the runner:

python -m experiments.our_decart.runner --dataset mnist --record-dim 784 --model-source trained --trained-models-dir experiments/models/trained

7. Baseline Comparison Workflow

For fair comparison, use a unified parameter set across schemes:

• N=10000
• n=32
• num_records=32
• record_dim=32
• policy_size=32
• num_runs=3
• dataset=synthetic

7.1 Run DeCart and DeCart* First

python -m experiments.our_decart.runner --dataset synthetic --N 10000 --n 32 --num-records 32 --record-dim 32 --policy-size 32 --num-runs 3
python -m experiments.our_decart_star.runner --dataset synthetic --N 10000 --n 32 --num-records 32 --record-dim 32 --policy-size 32 --num-runs 3

7.2 Run Baseline Schemes

python -m experiments.scheme1_ccs23.runner --dataset synthetic --N 10000 --n 32 --num-records 32 --record-dim 32 --policy-size 32 --num-runs 3
python -m experiments.scheme2_server.runner --dataset synthetic --N 10000 --n 32 --num-records 32 --record-dim 32 --policy-size 32 --num-runs 3
python -m experiments.scheme3_offline.runner --dataset synthetic --N 10000 --n 32 --num-records 32 --record-dim 32 --policy-size 32 --num-runs 3
python -m experiments.secpq.runner --dataset synthetic --N 10000 --n 32 --num-records 32 --record-dim 32 --policy-size 32 --num-runs 3 --model-types decision_tree

8. Revoke Experiments

python -m experiments.revoke.runner --scheme decart --model-type decision_tree --N 10000 --n 32 --num-records 10 --record-dim 10 --policy-size 32 --num-runs 3
python -m experiments.revoke.runner --scheme decart_star --model-type decision_tree --N 10000 --n 32 --num-records 10 --record-dim 10 --policy-size 32 --num-runs 3

9. Result Output Directories

By default, experiment results are stored in scheme-specific subdirectories under experiments/results/ (or override with --results-dir).

Typical directories:

• experiments/results/our_decart/
• experiments/results/our_decart_star/
• experiments/results/scheme1_ccs23/
• experiments/results/scheme2_server/
• experiments/results/scheme3_offline/
• experiments/results/secpq/
• experiments/results/revoke/

Plot-generation scripts are located in the image output directory experiments/results/pic_new/ and can be run directly from the project root:

python experiments/results/pic_new/communication/generate_communication_charts.py
python experiments/results/pic_new/computation/generate_computation_charts.py
python experiments/results/pic_new/size/generate_size_charts.py

Corresponding output directories:

• experiments/results/pic_new/communication/
• experiments/results/pic_new/computation/
• experiments/results/pic_new/size/

10. Windows Notes

• It is recommended to always run experiments with Python from your virtual environment.
• If PowerShell output redirection causes encoding errors, set:

$env:PYTHONIOENCODING = "utf-8"

11. Suggested Reproduction Flow

1. Create a virtual environment and install dependencies.
2. Run tests/test_schemes_comparison.py as a basic sanity check.
3. Run small-scale checks for our_decart and our_decart_star first.
4. Run other baseline schemes and revoke experiments.
5. Finally, run the plotting scripts to generate paper figures.