forgetnet

🛡️ ForgetNet: Differentially Private Block-wise Gradient Shuffle for Deep Learning 🧠

ForgetNet introduces a novel privacy-preserving technique for deep learning: Differentially Private Block-wise Gradient Shuffle (DP-BloGS). 🔒🔀

🌟 Features

• 🚀 Fast training times, close to non-private training
• 🎯 Competitive privacy guarantees compared to DP-SGD
• 📊 Better privacy-utility trade-off in many scenarios
• 🏋️ Scalable to large models (tested up to 1.1 billion parameters)
• 🧮 Parameter-wise privacy budget allocation

📦 Installation

pip install forgetnet

🚀 Quick Start BloGSSFTTrainer

from forgetnet import BloGSSFTTrainer
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load your model and tokenizer
model = AutoModelForCausalLM.from_pretrained("gpt2")
tokenizer = AutoTokenizer.from_pretrained("gpt2")

# Initialize the DP-BloGS trainer
trainer = BloGSSFTTrainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    tokenizer=tokenizer,
    dataset_text_field="text",
    target_epsilon=1.0,
    delta=1e-5,
    clip_value=1.0
)

# Train your model with privacy guarantees
trainer.train()

🚀 Quick Start with Privacy Engine and ResNet

Here's how to use the BloGS Privacy Engine with a ResNet model for MNIST classification:

import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import datasets, transforms
from torchvision.models import resnet18
from torch.utils.data import DataLoader
from forgetnet import BloGSPrivacyEngine

# Modify ResNet18 for MNIST (1 channel input instead of 3)
def mnist_resnet18():
    model = resnet18()
    model.conv1 = nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3, bias=False)
    model.fc = nn.Linear(model.fc.in_features, 10)  # 10 classes for MNIST
    return model

# Hyperparameters
batch_size = 64
learning_rate = 0.01
epochs = 10
target_epsilon = 1.0
delta = 1e-5
clip_value = 1.0

# Device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load MNIST dataset
transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.1307,), (0.3081,))
])
train_dataset = datasets.MNIST('./data', train=True, download=True, transform=transform)
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)

# Calculate total_iterations based on dataset size
total_iterations = (len(train_dataset) // batch_size) * epochs

# Initialize model
model = mnist_resnet18().to(device)

# Initialize optimizer
optimizer = optim.SGD(model.parameters(), lr=learning_rate)

# Wrap the optimizer with the PrivacyEngine
privacy_engine = BloGSPrivacyEngine(
    optimizer=optimizer,
    model=model,
    target_epsilon=target_epsilon,
    delta=delta,
    clip_value=clip_value,
    steps=total_iterations,
    batch_size=batch_size
)

# Training loop
model.train()
for epoch in range(epochs):
    for batch_idx, (data, target) in enumerate(train_loader):
        data, target = data.to(device), target.to(device)
        privacy_engine.zero_grad()
        output = model(data)
        loss = nn.functional.cross_entropy(output, target)
        loss.backward()

        epsilon_spent, delta = privacy_engine.step()

        if batch_idx % 100 == 0:
            print(f'Epoch {epoch}, Batch {batch_idx}, Loss: {loss.item():.4f}, Epsilon: {epsilon_spent:.4f}')

# Get total privacy spent after training
total_epsilon_spent = privacy_engine.get_privacy_spent()
print(f"Total privacy spent: ε = {total_epsilon_spent:.4f}")

This example demonstrates:

• Setting up a ResNet18 model modified for MNIST
• Loading the MNIST dataset
• Initializing the PrivacyEngine with the model and optimizer
• Training the model with privacy-preserving gradient updates
• Monitoring privacy budget expenditure during training

Adjust hyperparameters as needed for your specific use case.

📚 How It Works

DP-BloGS introduces a probabilistic approach to gradient noise through block-wise shuffling:

• 📊 Divide gradients into blocks
• 🔀 Shuffle blocks randomly
• 📏 Apply layer-specific block sizes
• ✂️ Use batch layer clipping
• 🧮 Accumulate gradients

This combination allows for fast training while maintaining strong privacy guarantees!

📈 Performance

DP-BloGS has been tested on various model architectures, including:

• GPT-2 (124M)
• BERT (110M)
• OPT (350M)
• BLOOM (560M)
• TinyLlama (1.1B)

Results show competitive or better performance compared to DP-SGD in terms of:

• 🏃‍♂️ Training speed
• 🎭 Privacy guarantees
• 📊 Model utility

🕵️ Membership Inference Attack (MIA)

ForgetNet now includes a powerful Membership Inference Attack tool to assess the privacy risks of your language models:

🚀 Quick Start

from forgetnet import LanguageMIA

# Initialize the MIA tool
mia = LanguageMIA()

# Perform the attack
results = mia.attack(train_dataset, test_dataset, model, tokenizer)

# Print the results
print(f"ROC AUC: {results['roc_auc']:.4f}")
print(f"Precision-Recall AUC: {results['precision_recall_auc']:.4f}")
print(f"Best model: {results['best_model']}")
print(f"Optimal threshold: {results['optimal_threshold']:.4f}")

📊 Comprehensive Evaluation

The LanguageMIA class provides a detailed analysis of your model's vulnerability to membership inference attacks:

• 🎯 ROC AUC: Measures the overall performance of the attack
• 📈 Precision-Recall AUC: Assesses the trade-off between precision and recall
• 🧠 Best Model: Identifies the most effective attack model
• 🔍 Optimal Threshold: Determines the best decision threshold for classification
• 📉 Standard and Optimal Metrics: Provides accuracy, precision, recall, and F1 score for both standard (0.5) and optimal thresholds

🛠️ Integration with Model Evaluation

Easily incorporate MIA into your model evaluation pipeline:

def evaluate_model(model, train_dataset, test_dataset, tokenizer):
    # Perform membership inference attack
    mia = LanguageMIA()
    mia_results = mia.attack(train_dataset, test_dataset, model, tokenizer)

    # Evaluate perplexity (assuming a function evaluate_perplexity exists)
    perplexity = evaluate_perplexity(model, test_dataset, tokenizer)

    # Combine results
    results = {
        'perplexity': perplexity,
        'mia_results': mia_results,
    }

    return results

# Usage
evaluation_results = evaluate_model(model, train_dataset, test_dataset, tokenizer)
print(f"Model Perplexity: {evaluation_results['perplexity']:.2f}")
print(f"MIA ROC-AUC: {evaluation_results['mia_results']['roc_auc']:.4f}")

🎯 Benefits

• 🔒 Quantify Privacy Risks: Understand your model's vulnerability to membership inference attacks
• 📈 Track Improvements: Monitor how privacy-preserving techniques affect model privacy
• 🚀 Optimize Trade-offs: Fine-tune the balance between utility and privacy in your models

Use the LanguageMIA tool to ensure your language models are both powerful and privacy-preserving!

📄 Citation

If you use ForgetNet in your research, please cite my paper:

@article{zagardo2024dpblogs,
  title={Differentially Private Block-wise Gradient Shuffle for Deep Learning},
  author={Zagardo, David},
  journal={arXiv preprint arXiv:2407.21347},
  year={2024},
  note={arXiv:2407.21347 [cs.LG]}
}

🤝 Contributing

We welcome contributions! Please see my CONTRIBUTING.md for details on how to get started.

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgements

We thank the open-source community and the authors of the papers cited in our work for their valuable contributions to the field of privacy-preserving machine learning.

Built with 🧠 by David Zagardo