DS-NSCA: Differentiable Search for Non-Profiling Side-channel Analysis

This repository contains the implementation of the DS-NSCA algorithm, along with supplementary materials for generating additional figures and reproducing the results reported in the paper.

Getting started

To avoid any conflicts, we recommend using a virtual environment. The code can be downloaded from our anonymous GitHub repository. Create a new Conda environment :

conda create -n ds_nsca python=3.8

Activate the environment:

conda activate ds_nsca

Install the required packages. We provide a requirements.txt file that lists all necessary dependencies:

pip install -r requirements.txt

Requirements

Here is a non-exhaustive list of the required packages with their default versions provided in the requirements.txt file:

• torch (1.12.1)
• numpy (1.23.4)
• h5py (3.7.0)

Additionally, the CUDA version on our platform is 12.0.

Configurations of DS-NSCA

Hyperparameters

The different hyperparameters of DS-NSCA are:

Hyperparameter	Description	Default	Suggested Range
train_loss_fn	Loss function for training model parameters $$\theta$$	MSE	-
val_loss_fn	Loss function for validation to evaluate model performance	Pearson correlation coefficient	-
epochs	Total number of training epochs	250	50–300, depending on dataset size and convergence
scaling coefficient $$\lambda$$	Weight of validation feedback in multi-objective training	0.25	0–0.5 for stable performance with limited data
decay_factor	Factor controlling exponential decay of learning rates or temperature schedules	0.97	0.90–0.99, balancing stability and convergence speed
key_lr	Learning rate for key probability parameters $$\alpha$$	$$1 \times 10^{-2}$$	$$1 \times 10^{-3}$$ – $$1 \times 10^{-2}$$ for stable convergence
model_lr	Learning rate for model parameters $$\theta$$	$$1 \times 10^{-3}$$	$$1 \times 10^{-3}$$ – $$1 \times 10^{-2}$$, task-dependent
optimizer	Algorithm used to update parameters during training	Adam	-
warmup_ period	Number of epochs before applying validation feedback or other scheduled mechanisms	15	10–25, depending on dataset size and early prediction reliability

Model Architecture

For attack scenarios involving aligned traces, we recommend using an MLP model, as it has lower computational cost. For non-aligned traces, a CNN model is preferred. Compared with single-output methods, the multi-output (MOC) approach introduces additional network parameters to model multiple outputs effectively.

Implementations of these models are provided in the Until/Model.py.

Dataset

All datasets used in this work are publicly available. We do not redistribute the raw data and only provide access links below. Since some official links are no longer accessible, alternative links are provided where necessary.

• ASCAD: https://github.com/ANSSI-FR/ASCAD
• CHES_CTF: https://drive.google.com/file/d/1c6kHrBAn35w4a8TFWHu5QSPxKp5fM2Dc/view?usp=drive_link
• AES_HD(Original Release): https://github.com/gabzai/Methodology-for-efficient-CNN-architectures-in-SCA/blob/master/AES_HD/AES_HD_dataset.zip
• AES_HD (Updated Version): https://drive.google.com/file/d/1VTLwbhwh9qx4ZIAdOK4tap2hNJ1o6vJS/view
• AES_RD: https://github.com/gabzai/Methodology-for-efficient-CNN-architectures-in-SCA/tree/master/AES_RD/AES_RD_dataset
• ASCADv2: https://github.com/ANSSI-FR/ASCAD

Users should download the datasets from the corresponding links and organize them according to their experimental setup.

Examples

To facilitate the reproduction of our experiments, we provide runnable examples for each dataset, configured with the same hyperparameters as reported in the paper.
Each example is organized under a directory named after the corresponding dataset.
For instance, the example for the AES_HD dataset is located in the AES_HD/ directory.

In addition, the Until/ directory contains the core implementation files used across all experiments, including:

• DataLoader.py: defines a lightweight PyTorch Dataset wrapper (CustomDataset) that encapsulates input traces and corresponding labels, enabling seamless integration with PyTorch DataLoader.
• Lossfunction.py: defines the training and validation loss functions used in the proposed method.
• Trainer.py: implements the training pipeline.

Together, these components enable reproduction of the experimental results and extension to other datasets.

License

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