Deep Learning Side Channel Attack Project

A Rust-based deep learning project for side-channel attack analysis using the Burn framework.

Project Structure

.
├── src/
│   ├── main.rs      # Main program logic and dataset loading
│   ├── data.rs      # Dataset structures and handling
│   ├── model.rs     # Neural network model definition
│   ├── training.rs  # Training pipeline implementation
│   ├── inference.rs # Inference functionality
│   └── scoop.rs     # Additional model configurations

Features

• HDF5 dataset loading for profiling and attack traces
• Neural network model with configurable architecture
• Hyperparameter optimization with random search
• Support for RMSprop optimizer with weight decay
• Training pipeline with customizable epochs and batch size
• GPU acceleration using WGPU backend

Requirements

• Rust (latest stable version)
• HDF5 library
• GPU with WGPU support

Installation

1. Clone the repository
2. Install dependencies:

cargo build

Usage

Training

The model can be trained using the default configuration:

cargo run --release

Optimisation of HyperParameters

If you want to use the opti fonction, it is preferable to save the model testing in a file Since every hyperparameter is print with the dbg macro, you can keep the graphic from the burn framework and also saving the test:

cargo run --release 2> back.log

Configuration

Key parameters can be adjusted in the source code:

• NBS: Number of samples (currently 100)
• NBE: Number of epochs (currently 200)
• Batch size (default: 64)
• Learning rate and optimizer parameters

Model Architecture

The neural network model is configurable with:

• Input size: 256 features
• Adjustable hidden layer size
• Support for both training and inference modes

Acknowledgments

• Big thanks to Nathan ROUSSELOT and Karine HEYDEMANN for this opportunity to discover deep learning.
• Thanks to Nathan ROUSSELOT for the code of SCOOP, the optimizer from his research paper.