AI for Code-Based Cryptography: Code Distinguisher

A pipeline that generates error‑correcting–code datasets and then trains a deep learning model to distinguish Goppa and other structured codes from Random codes.

Paper: AI for Code-based Cryptography

Table of Contents

0. Requirements
1. Quick Start
2. Directory Layout
3. Dataset Generation
4. Training
5. Environment & Dependencies
6. Citation
7. License

---

Requirements

• MacOS or Linux
• At least one NVIDIA GPU
• Python version >= 3.10
• Pytorch preinstalled
• SageMath, see Dependencies on how to install.

Quick Start

# 1. Create environment (Conda + uv pip)
$ conda create -n ai4code python=3.10 -y
$ conda activate ai4code

# 2. Install dependencies
$ pip install -I uv
$ uv pip install -r requirements.txt

If you don't have sage installed, you can install it by following Dependencies

Now you can run the project by first generating some data then training on it:

Generating 10k samples of binary Goppa codes of code length $n=32$, extension degree $m=5$, and polynomial degree $t=2$ and 10k random codes with the same dimension $k=n-mt$:

# 3. Generate Goppa and Random datasets (less than a minute on 8‑core laptop)
$ ./scripts/gen_dataset.sh

Now, let's launch the training on gpu=0:

# 4. Train on the Goppa dataset (GPU 0)
$ ./scripts/train.sh   # if gpu is not available, change device to cpu

You should get a final log after 1m08:

INFO - 07/07/25 19:43:56 - 0:01:08 - {'eval/0/accuracy': 0.995, 'eval/0/recall_1': 1.0, 'eval/0/recall_0': 0.990, 'eval/eval_loss': 0.030}
INFO - 07/07/25 19:43:56 - 0:01:08 - Finishing Training
✅  Training finished.

---

Directory Layout

.
├── scripts/
│   ├── gen_dataset.sh              # data pipeline (generate + collect)
│   ├── train.sh                    # training wrapper
│   ├── generate_data.py            # paralellizable script to generate a lot of data
│   └── collect_data.py             # assemble data generated by many workers into one H5 file.
├── data/                           # generated datasets land here
│   ├── dataset_goppa_32_H5/        # example of generated goppa dataset
│   └── dataset_random_32_H5/       # example of generated random dataset
├── train.py                        # main training entry‑point
├── src/                            # source code
│   ├── trainer.py                  # Trainer 
│   ├── data/                       # contains Dataset, Tokenizer, Generator, and DataSource classes.
│   ├── model/                      # different models depending on task
│   └── ...                         # optim, logger, metrics, etc.
├── README.md                       # this file
├── requirements.txt                # environment dependencies
└── notebooks/

---

Dataset Generation

Key parameters are exposed at the top of the Bash script ./scripts/gen_dataset.sh:

NUM_WORKERS=10   # parallel processes
N_SAMPLES=10000  # per‑code family
CODE_LEN=32      # code length
T_ALT=2          # error‑correction capability
M_ALT=5          # extension degree

Modify them in one place and rerun.

---

Under the hood of those scripts, these are the commands being run for data generation:

1. scripts.generate_data for Goppa codes
2. scripts.collect_data on the generated shard
3. scripts.generate_data for Random codes
4. scripts.collect_data on the second shard

Spawn 10 workers to generate Goppa codes

$ python -m scripts.generate_data --code goppa --num_workers 10  --n_samples 10000 --code_len 32 --t_alt 2 --save_every 1000 --m_alt 5 --dump_path ./data --exp_name dataset_goppa_32

Gather generated data and store in H5 file.

$ python -m scripts.collect_data --data_path ./data/dataset_goppa_32 --n_samples 10000 --code goppa --code_len 32  --m_alt 5 --t_alt 2 --Q 2

Spawn 10 workers to generate Random codes

$ python -m scripts.generate_data --code random --num_workers 10  --n_samples 10000 --code_len 32 --t_alt 2 --save_every 1000 --m_alt 5 --dump_path ./data --exp_name dataset_random_32

Gather Random code data and store in H5 file.

$ python -m scripts.collect_data --data_path ./data/dataset_random_32 --n_samples 10000 --code random --code_len 32  --m_alt 5 --t_alt 2 --Q 2

---

Training

./scripts/train.sh starts distinguisher model training on GPU 0 by default.

All hyper‑parameters (batch sizes, validation cadence, etc.) are defined at the top of the script so you can version‑control them easily.

Generator Matrix Distinguisher: DeepDistinguisher

Ro run DeepDistinguisher on Goppa codes (versus Random codes by default), use task = code-dist-goppa and provide the data_path for goppa codes. (The data path for random codes should be at data_path.replace('goppa', 'random'))

$ CUDA_VISIBLE_DEVICES=0 python train.py --task 'code-dist-goppa'  --train_batch_size 32 --val_batch_size 1000 --eval_samples 1000 --train_samples 20000  --val_every 500 --log_every 10 --code_len 32 --data_path './data/dataset_goppa_32_H5/goppa_nmt_32_5_2/dataset_10K.h5' --m_alt 5 --t_alt 2  --tqdm True --Q 2

Generator Matrix Completion: DeepRecover

To run DeepRecover on goppa dataset, we use task=code-complete-goppa and specify the mask size n_masked which is how many matrix entries are hidden from the model. This argument can either be an integer $\geq 1$ or a float $ < 1$ in which case it's interpreted as the probability of hiding an entry.

$ CUDA_VISIBLE_DEVICES=0 python train.py --task 'code-complete-goppa'  --train_batch_size 32 --val_batch_size 1000 --eval_samples 1000 --train_samples 10000  --val_every 500 --log_every 10 --code_len 32 --data_path './data/dataset_goppa_32_H5/goppa_nmt_32_5_2/dataset_10K.h5' --m_alt 5 --t_alt 2  --tqdm True --Q 2 --n_masked 20

---

Environment & Dependencies

Installing SageMath (non‑pip dependency)

The project imports the sage module. Choose one of the methods below and make sure it is done before you run the generation script:

Platform	Command
Conda (Linux/macOS/Windows)	conda install -c conda-forge sage
macOS (Homebrew)	brew install sagemath
Ubuntu / Debian	sudo apt-get update && sudo apt-get install sagemath
Docker (isolated)	docker run --rm -it -v $PWD:/work -w /work sagemath/sagemath sage -python train.py ...

Test your install:

sage -python - <<'PY'
import sageall, sys
print("✅ SageMath available (Python %s)" % sys.version.split()[0])
PY

---

Citation

If you use this benchmark in your research, please use the following BibTeX entry.

@misc{cryptoeprint:2025/440,
      author = {Mohamed Malhou and Ludovic Perret and Kristin Lauter},
      title = {{AI} for Code-based Cryptography},
      howpublished = {Cryptology {ePrint} Archive, Paper 2025/440},
      year = {2025},
      url = {https://eprint.iacr.org/2025/440}
}

---

License

This code is made available under CC-by-NC, however you may have other legal obligations that govern your use of other content, such as the terms of service for third-party models.