nnaes 0.0.1

A neural-network implementation of the AES cryptographic algorithm.

Neural AES (NNAES)

A PyTorch implementation of the AES cryptographic algorithm using Deep Neural Networks (DNNs), based on the concept of Deep Neural Cryptography.

---

Attribution

This project is adapted from the original work:

• Repository: https://github.com/DavidGerault/deep_neural_cryptography
• Paper: Deep Neural Cryptography – Gerault et al.
  https://eprint.iacr.org/2025/288.pdf

The original implementation was reorganized and extended for usability, packaging, and experimentation purposes.

GNU General Public License v3 (GPLv3)

---

Description

This library implements AES-128 as a deterministic neural network, without any training process.

The model operates on binary inputs represented as tensors of shape:

(batch_size, 128)

Each AES block is encoded as a 128-dimensional binary vector.

Key characteristics

• Exact implementation of AES operations using neural networks
• No learning or training required (fully deterministic)
• Based on:
  • Linear layers + ReLU activations
  • Corner functions for Boolean logic
  • Fixed, analytically constructed weights

Input / Output utilities

• integer_to_bitvector: converts a 128-bit integer into a binary tensor
• state_as_ints: converts model outputs back to integers

---

What was done in this project

Compared to the original implementation, this project introduces:

• A clean and reusable Python package structure
• Improved modularity and readability
• Integration with PyTorch workflows
• Benchmarking tools for performance analysis
• Unit tests based on AES standard test vectors (FIPS / AESAVS)

---

Installation

Clone the repository and install locally:

pip install -e .

For development (tests, etc.):

pip install -e .[dev]

---

Usage

Basic example (encryption protected)

from nnaes import NeuralAES
from nnaes.utils import integer_to_bitvector, state_as_ints
import torch

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

key = 0x2b7e151628aed2a6abf7158809cf4f3c
plaintext = 0x3243f6a8885a308d313198a2e0370734

model = NeuralAES(
    secret_key=key,
    direction="Encryption",
    protected=True,
    epsilon=1/4
).to(device)

model.eval()

x = torch.tensor(
    [integer_to_bitvector(plaintext)],
    dtype=torch.float16
).to(device)

with torch.inference_mode():
    y = model(x)

ciphertext = state_as_ints(y)[0]
print(f"{ciphertext:032x}")

Expected output:

3925841d02dc09fbdc118597196a0b32

---

Notes

• The implementation uses torch.float16 by default for performance reasons.
• GPU execution is recommended for efficient benchmarking.
• On CPU, using float32 may improve numerical stability.

---

Disclaimer

This project is intended for research and educational purposes only. It is not optimized for production cryptographic use.