MLEncrypt-Research

Usage

• single: runs a single instance of neural key exchange
• multiple: runs many instances of neural key exchange, all with the same configuration (this can be useful for benchmarking)
• hparams: runs hyperparameter optimization (different instances of neural key exchange will likely have different configurations)

Note: It may appear that enabling JIT compilation results in a slowdown. If you don't want to run with JIT, then don't pass the XLA flags to the run command.

Note: Our script features very verbose logging to TensorBoard. Enable this by passing -tb to the run command.

Note: You can choose to not calculate the synchronization score for each iteration. Enable this by passing -b to the run command.

CPU

TF_XLA_FLAGS="--tf_xla_auto_jit=2 --tf_xla_cpu_global_jit" poetry run mlencrypt-research single

GPU

TF_XLA_FLAGS=--tf_xla_auto_jit=2 poetry run mlencrypt-research single

Acknowledgements

• https://github.com/farizrahman4u/neuralkey
  • Mr. Fariz Rahman's implementation was the initial basis for ours
• https://github.com/brigan/NeuralCryptography
  • ported Dr. Luís F. Seoane's C++ implementation of the probabilistic attack to TensorFlow