This repository contains the source code for the paper "Is MPC Secure? Leveraging Neural Network Classifiers to Detect Data Leakage Vulnerabilities in MPC Implementations", which has been accepted for publication at IEEE Symposium on Security and Privacy (S&P) 2025.
MPCGuard is a practical framework designed to detect data leakage vulnerabilities in multi-party computation (MPC) implementations. With the increasing adoption of MPC for privacy-preserving computations, ensuring their security against data leakage vulnerabilities is critical. Unlike traditional security analysis methods that focus on theoretical proofs, MPCGuard employs neural network classifiers to analyze real-world MPC implementations and identify potential vulnerabilities.
- Automated Data Leakage Detection: Utilizes neural network classifiers designed according to MPC characteristics to identify potential data leakage vulnerabilities.
- Delta-Based Vulnerability Localization: Efficiently pinpoints the source of leakage within the code.
- Docker
- CUDA 12.4 (for GPU acceleration, optional)
- NVIDIA GeForce RTX 3080 (or any CUDA-compatible GPU, optional)
Using GPU
sudo apt-get install nvidia-container-toolkit
sudo systemctl restart docker
git clone https://github.com/FudanMPL/MPCGuard.git
docker build -t mpcguard .
docker run --gpus all -it mpcguard bash
Without using GPU
git clone https://github.com/FudanMPL/MPCGuard.git
docker build -t mpcguard .
docker run -it mpcguard bash
MPCGuard currently supports:
- Crypten
- TF-Encrypted
- MP-SPDZ
You can test the following protocol without modifying any code:
mul(Multiplication)linear(Linear Combination)ltz(Less Than Zero)eqz(Equal to Zero)truncpr(Probabilistic Truncation)
Step 1.
cd /MPCGuard/frameworks/CrypTen
Step 2. Create a yaml file (e.g. test.yaml) with the following content
real_program: ["python", "./test_ass_eq.py"]
ideal_program: F_ass_eq # Choice: "F_ass_mul", "F_ass_eq", "F_ass_linear", "F_ass_ltz", "F_ass_truncpr", "F_rss_mul", "F_rss_linear", "F_rss_ltz", "F_rss_eq", "F_rss_truncpr"
party_number: 2
corrupted_party:
- 0
victim_party: 1
protocol_execution_times: 2000
accuracy_verify_gap: 0.1
use_model: MPCNN # Choice: MPCNN, BasicDNN, BasicCNN, BasicLSTM, xgboost
run_mode: Detect # Choice: Detect (if view exist, do not collect again), All (always re-collect the view)
bug_file: found_bugs_with_different_model/MPCNN/ass_eq
real_view_data_dir: real_view_data/ass_eq/
epochs: 200
Step 3.
python ../../main.py --config=./test.yaml
cd /MPCGuard
./test_with_different_models.shThe found bugs will be stored in /MPCGuard/frameworks/Crypten/found_bugs_with_different_model/, /MPCGuard/frameworks/TF-Encrypted/found_bugs_with_different_model/, and /MPCGuard/frameworks/MP-SPDZ/found_bugs_with_different_model/
cd /MPCGuard
./test_with_different_thresholds.shThe found bugs will be stored in /MPCGuard/frameworks/Crypten/found_bugs_with_different_threshold/, /MPCGuard/frameworks/TF-Encrypted/found_bugs_with_different_threshold/, and /MPCGuard/frameworks/MP-SPDZ/found_bugs_with_different_threshold/
cd /MPCGuard
./test_with_different_paddings.sh
The found bugs will be stored in /MPCGuard/frameworks/Crypten/found_bugs_with_different_padding/, /MPCGuard/frameworks/TF-Encrypted/found_bugs_with_different_padding/, and /MPCGuard/frameworks/MP-SPDZ/found_bugs_with_different_padding/
Step 1. Place your implementation in /MPCGuard/frameworks/[framework_name]/, and
cd /MPCGuard/frameworks/[framework_name]/
Step 2. Write a python program (e.g. test_new_protocol.py) that accepts the following arguments:
--try_times # Number of executions to collect views
--secret # Input secret value
--no-record # Skip recording views/inputs/outputs
--vul_index # Minimum adversary view length for leakage detection
--bug_file # Path to save bug info
--real_view_data_dir # Path to save collected views
In this step, you may need to manually instrument the code to collect the parties’ views, inputs, and outputs. See MPCGuard/frameworks/CrypTen/test_ass_ltz.py for reference.
Step 3. Create the corresponding functionality in /MPCGuard/functionality.py
For example,
def F_ass_linear(secret, a=1, b=2, c=3):
shares_x = generate_ass_shares(secret)
# Convert constants and reconstruct x and y from shares using numpy arrays
a = np.array(a, dtype=np.int64)
b = np.array(b, dtype=np.int64)
c = np.array(c, dtype=np.int64)
x = sum(np.array(shares_x, dtype=np.int64))
shares_y = [1 for i in range(len(shares_x))]
y = sum(np.array(shares_y, dtype=np.int64))
# Compute z = a*x + b*y + c
z = a * x + b * y + c
shares_z = generate_ass_shares(z)
ass_functionality_record(secret, shares_x, shares_z)
Step 4. Create a yaml file (e.g. test.yaml) with the following content
real_program: ["python", "./test_new_protocol.py"]
ideal_program: F_ass_linear
party_number: 2
corrupted_party:
- 0
victim_party: 1
protocol_execution_times: 2000
accuracy_verify_gap: 0.1
use_model: MPCNN # Choice: MPCNN, BasicDNN, BasicCNN, BasicLSTM, xgboost
run_mode: Detect # Choice: Detect (if view exist, do not collect again), All (always re-collect the view)
bug_file: found_bugs_with_different_model/MPCNN/ass_eq
real_view_data_dir: real_view_data/ass_eq/
epochs: 200
Step 5. Run
python ../../main.py --config=./test.yaml
This project is licensed under the MIT License. See the LICENSE file for details.
For any questions or contributions, please open an issue or submit a pull request!