Secret-key generation from wireless channels using variational autoencoders and domain adversarial neural networks

Abstract

Physical layer key distribution is an emerging technology in which legitimate communicating parties can extract secret keys with low computational cost and high security by using inherent randomness in the wireless channel. Due to a large number of wireless communication devices, the wireless channel has become weakly correlated with substantial noise, making the channel imperfect, leading to unreliable extraction of secret keys and increased key disagreement rates. However, existing knowledge is working for specific environments where the training and testing conditions are the same, but in the real world the environment is frequently changing, which existing knowledge did not take into account. To address these problems, we present a novel physical layer key distribution scheme based on variational autoencoders and domain adversarial neural network (VAEDANN) that ensure secret key extraction between legitimate devices from weakly correlated wireless channels. The variational autoencoder extracts the latent feature and eliminates the noise in the data, domain adversarial learning makes the model domain invariant. We used a quantization method called channel quantization alternating to address high key disagreement rates. To control information leakage and correct the discrepancy, we used cascade protocol and the toeplitz algorithm. After extensive simulation experiments, the VAEDANN-based method demonstrates a lower MSE compared to other methods across environments, highlighting its superior capability in learning channel reciprocity. Additionally, it achieves a lower bit disagreement rate and key error rate, exhibits high randomness in the key, and maintains a low computational cost. These features make the VAEDANN method highly reliable for resource constrained devices.