SecreatAI

SecreatAI is a game prototype built around AI-encoded file transfer. The goal is to train an autoencoder, use that model to transform a file into a .aiz file protected by a mask password, then challenge another player to reconstruct the file with their own model or their own analysis methods.

Important: this project is a playful and educational experiment about autoencoders, binary representations, and pattern analysis. It is not a reliable encryption tool for protecting sensitive data.

Game Objective

Each player has an autoencoder model trained separately. One player encodes a file with their model, shares only the .aiz file, and the opponent tries to decode it.

The encoding player scores points if the opponent cannot reconstruct the file. The opponent scores points if they manage to produce a correct reconstruction.

Features

• Tkinter GUI for training, encoding, and decoding.
• Train or resume training of a .pkl model.
• Encode files into .aiz with a random salt and mask password.
• Decode files into the decoded/ folder.
• Training loss curve in the interface.
• Pattern analysis tool for comparing two .aiz files.
• Separate scripts for training, encoding, and decoding.

Requirements

• Python 3.10, configured by .python-version.
• pip.
• tkinter for the graphical interface.

On some Linux distributions, tkinter must be installed through the system package manager, for example:

sudo apt install python3-tk

Installation

From the project folder:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
export MPLCONFIGDIR=.matplotlib-cache

The current Python dependencies are intentionally lightweight:

• numpy
• matplotlib

The project also uses modules from the Python standard library, including pickle, lzma, hashlib, tarfile, threading, and tkinter.

Quick Start

The graphical interface is the easiest way to use the project:

python SecreatAI_GUI.py

In the interface:

1. Choose an existing model or click New Model.
2. Click Train / Resume to create or continue training the model.
3. Choose a file with Browse.
4. Click Encode, enter a password, then produce a .aiz file.
5. Click Decode, enter the same password, then reconstruct a file from a .aiz.

Decoded files are written to decoded/. In the interface, files chosen outside the project folder are first copied into the application folder before encoding or decoding.

Script Usage

The main scripts can also be called from Python.

Train a Model

python Training_only.py

By default, this script creates or resumes model.pkl.

To use another model name:

python -c "import Training_only; Training_only.model_name='alice.pkl'; Training_only.main()"

Encode a File

python -c "import Encode_only; Encode_only.main('model.pkl', 'mon_fichier.pdf', mask_password='secret')"

If the internal reconstruction reaches 100%, a mon_fichier.pdf.aiz file is created in the same folder as the source file, unless an output folder is provided.

Example with an output folder:

python -c "import Encode_only; Encode_only.main('model.pkl', 'mon_fichier.pdf', output_dir='encoded', mask_password='secret')"

If mask_password is not provided, the script asks for the password in the terminal.

Decode a .aiz File

python -c "import Decode_only; Decode_only.main('model.pkl', 'mon_fichier.pdf.aiz', mask_password='secret')"

By default, the result is written to decoded/ with the .aiz suffix removed. For .aiz files produced with a password, the same password is required for decoding.

Compare Two .aiz Files

python pattern_aiz_compare.py fichier1.aiz fichier2.aiz

This script displays statistics about 64-bit blocks, repetitions, entropy, and Hamming distance. It helps identify whether visible patterns remain in encoded files.

Proposed Rules

1. Each player trains their own .pkl model.
2. The active player chooses a source file.
3. They encode the file into .aiz and verify that their own model can reconstruct the file.
4. They keep their password secret or define a sharing rule before the round.
5. They send only the .aiz file to the opponent.
6. The opponent tries to decode the file with their model or with any other analysis strategy.
7. The original file is revealed.
8. Points are awarded:
   • opponent wins if the file is correctly reconstructed;
   • encoder wins if the file remains undecipherable.
9. Roles rotate in the next round.

File Structure

.
|-- SecreatAI_GUI.py                 # Graphical interface
|-- Training_only.py                 # Train/resume a model
|-- Encode_only.py                   # Encode a file to .aiz
|-- Decode_only.py                   # Decode a .aiz
|-- Autoencoder_Encoder_Decoder.py   # Historical combined version
|-- Autoencoder_Encoder_Decoder_timed.py
|-- pattern_aiz_compare.py           # Pattern analysis between .aiz files
|-- Compare_files.py                 # Simple binary file comparison
|-- requirements.txt
|-- model.pkl                        # Default model, if present
`-- decoded/                         # Decoded outputs

Generated Files

• *.pkl: trained models. The provided model.pkl file is used as the default model.
• *.aiz: encoded files. These outputs are generated locally and ignored by Git.
• decoded/: reconstructed files.
• .matplotlib-cache/: local Matplotlib cache.
• __pycache__/: Python cache.

These files are ignored by .gitignore when they are temporary or generated.

Technical Notes

The model works on 8-bit input blocks and produces 64-bit encoded representations. The original file bytes are first masked with a random salt and optional password, then the masked bytes are encoded by the model and compressed with lzma before the .aiz file is written. If a password is provided, it is combined with the salt using sha256 to produce the mask stream.

Training uses a custom dense neural network implementation and the Adam optimizer with numpy. Training and validation losses are stored in the .pkl file, allowing the interface to display the history.

Known Limitations

• The .aiz format is not cryptographic encryption.
• The mask password makes analysis harder, but it does not replace real encryption.
• The same model can reveal exploitable regularities depending on the files and training level.
• The CLI scripts do not have an argparse interface yet; advanced calls are therefore made through python -c or through the GUI.
• pickle must not be used with untrusted models: loading a .pkl file from an unknown source can execute arbitrary code.

Troubleshooting

Matplotlib-Related Error

Define a local cache folder:

export MPLCONFIGDIR=.matplotlib-cache

No module named tkinter Error

Install the Tkinter system package matching your Python. On Debian/Ubuntu:

sudo apt install python3-tk

Decoding Produces an Incorrect File

Check that the .aiz was decoded with the same model and the same password used for encoding. In the game, using another model is part of the challenge, but exact reconstruction is then not guaranteed.

Invalid password or corrupted AIZ payload Error

The provided password does not match the .aiz file, or the encoded file is incomplete/corrupted. Try again with the password used during encoding.

License

This project is distributed under the MIT license. See LICENSE.