QuanCrypt-FL: Quantized Homomorphic Encryption with Pruning for Secure and Efficient Federated Learning

Overview

QuanCrypt-FL is an innovative approach designed to enhance security in federated learning by combining quantization and pruning strategies. This integration strengthens the defense against adversarial attacks while also lowering computational overhead during training. Additionally, the approach includes a mean-based clipping mechanism to address potential quantization overflows and errors. The combination of these techniques results in a communication-efficient FL framework that prioritizes privacy without significantly affecting model accuracy, thereby boosting computational performance and resilience to attacks.

Features

• Federated Learning (FL) setup with customizable client count and local training parameters.
• CKKS scheme-based homomorphic encryption for secure aggregation.
• Quantization for efficient communication and storage.
• Pruning to optimize the model and reduce communication overhead.
• Support for CIFAR-10, CIFAR-100, and MNIST datasets with IID and non-IID partitioning.
• Integration of early stopping to enhance training efficiency.
• Comprehensive logging and result storage in CSV files.
• Inverse federated attacks to test model robustness.
• Customizable training, quantization, and encryption parameters.

Installation

1. Clone the repository:

   git clone https://github.com/juealcs/QuanCryptFL.git
   cd QuanCryptFL

2. Install the required Python packages:

   pip install -r requirements.txt

3. Ensure you have PyTorch and TenSEAL installed. For installation details, refer to the PyTorch installation guide and TenSEAL documentation.

Requirements

• Python 3.7+
• PyTorch
• TenSEAL
• Torchvision
• CUDA (for GPU support)
• InverseFed

Arguments

QuanCryptFL supports a variety of arguments for customizing the FL setup, training, and evaluation. Below is a list of available arguments and their descriptions:

Argument Name	Description	Default Value
--device	Device to be used for Federated Learning (GPU or CPU)	GPU
--dataset	Dataset to use (MNIST, CIFAR10, CIFAR100)	CIFAR100
--partition	Data partitioning strategy (IID or nonIID)	IID
--data_dir	Directory for dataset storage	./data
--model_name	Model to use (CNNMNIST, CNNCIFAR, AlexNet, ResNet18)	ResNet18
--n_clients	Number of clients for federated learning	50
--n_local_epochs	Number of local epochs per client	1
--num_rounds	Number of communication rounds	300
--learning_rate	Learning rate for the optimizer	0.001
--weight_decay	Weight decay for the optimizer	0.0001
--batch_size	Batch size for client-side training	64
--num_bits	Number of bits for quantization	8
--clip_factor	Clip factor for model updates (1.0-5.0)	3.0
--initial_pruning_rate	Initial pruning rate	0.2
--target_pruning_rate	Target pruning rate	0.5

Usage

To run the Vanilla FL script with the desired configurations, use the following command format:

python Vanilla-FL.py --model_name ResNet18 --dataset CIFAR10 --n_clients 10 --n_local_epochs 1 --num_rounds 300 --batch_size 64 --learning_rate 0.001 --partition IID --lambda_param 0.1 --reconfig_freq 50 --patience 5 --perform_attack True

Run the QuanCryptFL mechanism use the following script with the desired configurations:

python QuanCrypt-FL.py --model_name ResNet18 --dataset CIFAR10 --n_clients 10 --n_local_epochs 1 --num_rounds 300 --batch_size 64 --learning_rate 0.001 --partition IID --clip_factor 0.5 --num_bits 8 --lambda_param 0.1 --reconfig_freq 50 --patience 5 --perform_attack True