International Association for Cryptologic Research

International Association
for Cryptologic Research

Transactions on Cryptographic Hardware and Embedded Systems 2026

Coil-Based Detection and Concurrent Error Correction Against EMFI


README

📊 Dataset EMFI Coil Evaluation

This repository contains the datasets and analysis scripts supporting the CHES 2026 publication:
"Coil-Based Detection and Concurrent Error Correction Against EMFI: An Experimental Case-Study on a Prototype ASIC" (eprint).
It enables the reproduction of all results and plots presented in the paper.

📋 Table of Contents

🛠️ Prerequisites

To reproduce the results, you need:

  1. Git LFS - Required for cloning large data files
  2. Environment setup (choose one):
    • Nix package manager (recommended for reproducibility)
    • Python 3.13+ with pip (tested with Python 3.13.9, see requirements.txt)

📂 Repository Structure

Dataset-EMFI-Coil-Evaluation/
├── implementations/           # Target implementations source code
|   ├── v3                     # Design files τ₁ (Unprotected AES)
|   ├── v4                     # Design files τ₂ (AGEFA AES [3, 1, 3]-code)
|   ├── v5                     # Design files τ₃ (AGEFA AES [16, 8, 5]-code)
|   ├── aes-byte-serial-tb.vhd # Testbench for all designs
├── measurements/              # Raw measurement data
│   ├── v3-10rep-1_1mx1_1mm/   # Target τ₁ (Unprotected AES)
│   │   ├── Y0_X0_Z0/          # Probe position (Y=0, X=0, Z=0)
│   │   └── ...
│   ├── v4-10rep-1_1mx1_1mm/   # Target τ₂ (AGEFA AES [3, 1, 3]-code)
│   ├── v5-10rep-1_1mx1_1mm/   # Target τ₃ (AGEFA AES [16, 8, 5]-code)
│   └── ...
├── results/                   # Generated CSV files and plots
├── read.py                    # Main script for data processing
├── shell.nix                  # Nix environment configuration
└── README.md

🚀 Reproducing Results

1. Install Git LFS (if not already installed)

# On Ubuntu/Debian
sudo apt-get install git-lfs

# On macOS
brew install git-lfs

# On other systems, see: https://git-lfs.github.com/

Initialize Git LFS:

git lfs install

2. Clone the Repository

git clone https://github.com/ChairImpSec/Dataset-EMFI-Coil-Evaluation
cd Dataset-EMFI-Coil-Evaluation

[!IMPORTANT]
Due to Git LFS budget limitations, it is currently not possible to clone the preprocessed cached files (.parquet files in results/).
The corresponding error can be ignored.
However, you must regenerate all temporary files using the make full-rebuild command (see section 4 below).
This will process the raw measurement data (~30 minutes) to generate all analysis results.

3. Set Up the Environment

Option A: Using Nix Flakes (recommended - provides pinned dependencies):

nix develop

For setups without experimental features enabled:

nix --extra-experimental-features 'nix-command flakes' develop

Option B: Using traditional Nix:

nix-shell

Option C: Using pip (without Nix):

pip install -r requirements.txt

4. Run the Analysis Script

Execute the following command to process cached data and generate CSV files:

python read.py

By default, this uses cached preprocessed data (included in the repository) and generates only pinout-based detection heatmaps. For full analysis from raw data, see the command-line options below.

Command-Line Options

The script supports the following command-line arguments (use python read.py --help to see all options):

Example usage:

# Quick start - use cached data with partial analysis (default)
python read.py

# Command equivalent to the original submitted version
python read.py --clean --all partial

# Full analysis - required to run the scripts in `results` (takes ~30 minutes)
python read.py --clean --no-show-plots --all all

# Regenerate with partial analysis without showing plots
python read.py --clean --no-show-plots

[!NOTE]
When all plots are generated, 4 of these plots will contain only zeros for all coordinates.
The reason for this is a bug in the HDL implementation of the ASIC hosting the physical and algorithmic countermeasures:
The four HVT registers with the highest index are not connected to the data bus correctly.

Experiments evaluated:
The script processes three target implementations: v3-10rep-1_1mx1_1mm, v4-10rep-1_1mx1_1mm, and v5-10rep-1_1mx1_1mm.

[!NOTE]
The script generates more plots than depicted in the publication.

📈 Dataset Description

Measurement Folders

The measurements directory contains subfolders for each target, as described below:

Folder Target Description
v3-10rep-1_1mx1_1mm τ₁ (Unprotected AES)
v4-10rep-1_1mx1_1mm τ₂ (AGEFA AES [3, 1, 3]-code)
v5-10rep-1_1mx1_1mm τ₃ (AGEFA AES [16, 8, 5]-code)

File Naming Convention

Each measurement folder contains subfolders named Yy_Xx_Z0, where:

Inside these subfolders, binary files follow the naming scheme:
Expr_i_Cc_Xx_Yy_Z0_Vv_Pp_LJDd.dat

💻 Implementation Files

All target implementations are provided as RTL code in the implementations directory, along with a testbench that is compatible with all designs.

📄 Output Files

The read.py script generates CSV files for each plot in the paper.
These files are saved in the results directory, mirroring the structure of the measurements folder.

[!NOTE]
Some plots (e.g., the first four line plots per target) are not directly used in the publication, but post-processed.

✨ Reproduction Workflow

Quick Start (Using Cached Data)

The repository includes preprocessed data for fast reproduction:

make              # Run post-processing with cached data (~1 minute)

Complete Reproduction (From Raw Data)

To regenerate everything from scratch:

make full-rebuild    # Clean and regenerate everything (~30 minutes)

Step-by-Step Workflow

make clean            # Remove all generated files
make analysis-clean   # Generate CSV files from raw measurements (~30 min)
make post-process     # Run post-processing scripts for merged plots

Available Make Targets

Manual Execution (Without Make)

If you prefer to run scripts manually:

# Step 1: Generate initial CSV files
python read.py --clean --no-show-plots --all all

# Step 2: Run post-processing scripts
cd results
python ../merge-coil-polarity-results.py
python ../merge-coil-results.py
python ../merge-coil-types-for-polarity.py
python ../merge-coil-polarity-first-reaction.py

[!NOTE]
For completeness, we provide the output of all scripts within the results folder.

📊 Generating Paper Figures

Currently, no python script is available to generate exactly the same plot as shown in Figure 7a and Figure 7b,
due to manually post-processing and computations from within TikZ.

Figure 7a (Voltage vs Polarity Counts)

Figure 7a data is generated automatically by the function plot_coil_counts_by_vlevel_and_polarity(df, export_dir, id="").

Generation pipeline:

  1. Generate initial data:

    python read.py --all partial

    This opens plots for each target and all three cell types.

  2. Merge the data:

    • merge-coil-polarity-results.py merges the polarity results
    • merge-coil-types-for-polarity.py further merges across cell types

  3. Final visualization:
    The resulting CSV is processed by a TikZ script that computes ratios by dividing the detection counts by the number of experiments conducted for each voltage level and polarity.

Figure 7b (Polarity Comparison)

Figure 7b requires manual post-processing after running merge-coil-polarity-first-reaction.py.

The function plot_polarity_start_comparison(df, export_dir, id="") generates 3 versions of Figure 7b (one per target).
The final Figure 7b in the paper is the average of these 3 plots, with coil 9 manually adjusted.

The file results/polarity-comparison-results-merged.csv contains the merged results.
After post-processing the result is as follows:

coil,value
coils-lvt-c0,0
coils-lvt-c1,0
coils-lvt-c2,3
coils-lvt-c3,0
coils-lvt-c4,3
coils-lvt-c5,3
coils-lvt-c6,3
coils-lvt-c7,0
coils-lvt-c8,3
coils-lvt-c9,1.5
coils-lvt-c10,3

📧 Contact

For questions or issues, please contact:
[email protected] or open an issue on the repository.