Learning to Protect Communications with Adversarial Neural Cryptography

TL;DR

This work demonstrates that neural networks can learn to protect communications from an adversarial neural observer through end-to-end adversarial training, without prescribing cryptographic algorithms. By modeling Alice, Bob, and Eve and optimizing a joint objective that rewards Bob’s accurate recovery of plaintext while hindering Eve, the authors show that the networks can discover encryption and selective protection schemes. The experiments across varying key lengths reveal robust, key-dependent enciphering behavior and reveal challenges in training stability and generalization, offering a proof-of-concept for data-driven cryptographic protection. The results motivate further exploration of neural approaches to cryptography, privacy-preserving representations, and related robustness analyses.

Abstract

We ask whether neural networks can learn to use secret keys to protect information from other neural networks. Specifically, we focus on ensuring confidentiality properties in a multiagent system, and we specify those properties in terms of an adversary. Thus, a system may consist of neural networks named Alice and Bob, and we aim to limit what a third neural network named Eve learns from eavesdropping on the communication between Alice and Bob. We do not prescribe specific cryptographic algorithms to these neural networks; instead, we train end-to-end, adversarially. We demonstrate that the neural networks can learn how to perform forms of encryption and decryption, and also how to apply these operations selectively in order to meet confidentiality goals.

Figures (6)

• Figure 1: Alice, Bob, and Eve, with a symmetric cryptosystem.
• Figure 2: Evolution of Bob's and Eve's reconstruction errors during training. Lines represent the mean error across a minibatch size of 4096.
• Figure 3: Final reconstruction errors of Bob and of the most effective retrained Eve, for the fourteen initially successful runs, of twenty. An ideal result would be a dot in the upper-left corner, representing no Bob reconstruction error and 8 bits of Eve reconstruction error.
• Figure 4: Training to estimate D while hiding C.
• Figure 5: Alice, Bob, and Eve, with an asymmetric cryptosystem.