Neural Cryptography

Project Aim

To build an encryption system using Neural Networks without specifying any particular algorithm.

Members

• Shreya Raghavendra
• Ankit Karody
• Anirudh Sujish
• Arjun Anil Pillai
• Hrithik Bhat
• Mansi
• Nihar Rai
• Mohammed Rushad
• Sumanth Reddy

Work Done

The model is trained adversarially to learn what information to encrypt and decrypt and how to do so. The model would consist of neural networks named Alice and Bob, limiting what a third neural network named Eve learns from eavesdropping on the communication between Alice and Bob.

1. Creating Individual Tensors
2. Adding relevant layers to each model
3. Adding the tanh activation function
4. Mimicking Symmetric Cryptosystem
5. Writing the Loss Functions for Alice and Bob, and Eve
6. Introducing the adversarial nature in the loss functions
7. Creating an appropriate random dataset for training
8. Culminating the functions worked on to train the NN.

Results