DAGFed: A Dual-Perturbation and Adaptive Graph Framework for Secure and Robust Distributed Learning

This repository contains the official implementation of DAG_Fed, a federated learning framework integrating Dynamic Adaptive Gradient Clipping (DAGC) and Hierarchical Differential Privacy (HDP) to improve both privacy protection and model utility under heterogeneous (non-IID) data distributions.

---

📌 Features

• Dynamic Client Selection & Adaptive Gradient Clipping — Mitigates statistical heterogeneity by adaptively tuning gradient norms per client.
• Hierarchical Differential Privacy (HDP) — Multi-level noise injection for balanced privacy–utility trade-off.
• Modular Aggregator — Supports FedAvg, SCAFFOLD, and custom aggregation strategies.
• Flexible Data Partitioning — IID and non-IID splits via Dirichlet sampling.
• Reproducible Experiments — Unified config file and logging system.

---

📂 Project Structure

DAGC_HDP_Federated_Learning/
│
├── config.py                 # Global configuration: hyperparameters, DP budget, scenarios
├── main.py                   # Main entry: training orchestration, global loop
│
├── clients/
│   ├── client.py             # Client base class: data loading, local training, DAGC
│   ├── dp_utils.py           # DP utilities: gradient clipping, noise injection
│   ├── noise_scheduler.py    # Dynamic noise scheduling
│
├── server/
│   ├── aggregator.py         # Aggregators (FedAvg, SCAFFOLD, etc.)
│   ├── privacy_accountant.py # DP budget tracking, RDP accounting
│
├── models/
│   ├── model_cnn.py          # CNN model for vision datasets (CIFAR-10, MNIST)
│
├── data/
│   ├── data_loader.py        # Dataset loading & non-IID partitioning
│   ├── datasets/             # Dataset storage
│
├── experiments/
│   ├── logger.py             # Logging utilities
│   ├── metrics.py            # Accuracy, F1, MSE calculation
│   ├── attack_test.py        # Privacy attack evaluation scripts
│
├── utils/
│   ├── helper.py             # General utilities (seed, save/load, etc.)
│   ├── plots.py              # Visualization of results
│
├── README.md                 # Project documentation
└── requirements.txt          # Python dependencies

---

⚙️ Installation

1. Clone the repository

git clone https://github.com/dxpython/DAG_Fed.git
cd DAG_Fed

2. Create a Python environment

conda create -n DAG_Fed python=3.9
conda activate DAG_Fed

---

🚀 Usage

Run a CIFAR-10 experiment with 10 clients, 100 rounds, and a privacy budget of ε = 1.0:

python main.py --dataset cifar10 --rounds 100 --clients 10 --epsilon 1.0

Key Arguments:

• --dataset : cifar10, cifar100, mnist, fashionmnist
• --rounds : Number of global communication rounds
• --clients : Number of participating clients
• --epsilon : Privacy budget for DP
• Additional parameters can be set in config.py

---