Gradient Inversion Attack: Deep Leakage from Gradients (DLG)

This repository implements the Deep Leakage from Gradients (DLG) attack, demonstrating how sensitive training data can be reconstructed from gradients shared during Federated Learning (FL).

---

🛠 Features

• Supports LeNet and ResNet-18 architectures for the attack on CIFAR-10.
• CLI-based workflow for easy experimentation.
• Generates intermediate and final reconstructed images for visualization.

---

🚀 Getting Started

Prerequisites

• Python 3.8 or higher
• PyTorch
• torchvision
• matplotlib

Installation

1. Clone the repository:

   git clone https://github.com/your-username/gradient-inversion-attack.git
   cd gradient-inversion-attack

2. Install dependencies:

   pip install -r requirements.txt

---

🔧 Running the Attack

Use main.py to run the attack via the command line. The script supports the following arguments:

CLI Arguments

Argument	Type	Default	Description
--model	string	None	Specify the model: lenet or resnet. (Required)
--img_idx	integer	25	Index of the CIFAR-10 image to attack.
--epochs	integer	500	Number of optimization iterations for the attack.

Example Command

To run the attack on LeNet for 500 epochs:

python main.py --model lenet --img_idx 25 --epochs 500

To switch to ResNet:

python main.py --model resnet --img_idx 30 --epochs 1000

---

📊 Outputs

1. Ground Truth Image: The original CIFAR-10 image selected for the attack, saved as gt_image.png in the plots/ directory.
2. Dummy Image Initialization: Randomly initialized dummy image (dummy_img.png).
3. Reconstructed Images: Saved progressively during the attack (reconstruction.png), visualizing the attack's performance.

---

🔍 Observations

• LeNet: Successfully reconstructs images in ~300-500 iterations, with results closely matching the ground truth.
• ResNet-18: Reconstruction is more challenging due to the complexity of the model and gradient dispersion but remains feasible under certain conditions.

---

🌱 Next Steps

• Test the attack on deeper models like ResNet-56.
• Experiment with mitigation strategies such as gradient perturbation and compression.
• Extend support for other datasets (e.g., MNIST, ImageNet).
• Understand what works and what doesn't when dealing with deeper models.

---

📚 References

• Original Paper: Deep Leakage from Gradients
• Project Insights: Implementation of the DLG attack on CIFAR-10 using LeNet and ResNet.

---

Star the repo if you find it helpful! 🌟