Learning asymmetric encryption using adversarial neural networks

We propose a multi-agent adversarial neural networks model where a sender (Alice) and a receiver (Bob) are able to learn to use a pair of public/private keys in order to protect their communication from one or more attackers/eavesdroppers. Existing work in the field required shared symmetric information between Alice and Bob before initiating the training process. To the best of our knowledge, this is the first work in which Alice and Bob with asymmetric information can train themselves to protect their communication. Our initial model setup contains five agents: sender Alice, receiver Bob, eavesdropper Eve and two neural networks (we call them public keys generator and private keys generator) that, based on a (secret) random noise from Bob, will generate a pair of public/private keys that allows Alice to encrypt a message with the public key and Bob to decrypt the message with the private key while preventing Eve from decrypting the secret message using the public key. We show that the neural networks are able to establish a communication and secure it from Eve. Finally, we consider adversaries stronger than Eve to model leakage attacks, chosen plaintext attacks (CPA) and test the distinguishability between ciphertexts. The last three experiments show that neural networks (with asymmetric information) can secure the communication providing stronger security guarantees and resilience to leakage attacks which may include leakage from the private key.