LearningToProtect

Implementation of Learning to Protect Communications with Adversarial Neural Cryptography in PyTorch

Caution: Details of implementation differ much from the original paper. I tried to use dense(fc) layers with relu activation, instead of dense+conv1d+sigmoid. See model/model.py.

Requirements

pip install -r requirements.txt (PyTorch, Numpy, TensorboardX)

Train

python trainer.py -c config/default.yaml -n [name of run]

• You may copy cp config/default.yaml config/config.yaml and change parameters (e.g. size of plain/key/cipher, depth of NN, …) to experiment with your own setting.

Tensorboard

tensorboard --logdir logs/

Results

Result with config/default.yaml, trained on GTX 1080 for 1 hour.

• Accuracy: (Green: Bob, Gray: Eve, Orange: Random guess)
• Loss: (Green: B, Gray: E, Magenta: AB)

License

Apache License v2.0