Attack on Machine Learning based Physical Layer Key Generation scheme

Abstract:
In Physical Layer Key Generation, two legitimate parties extract the same key from channel measurements. There exist classical as well as machine-learning based approaches. One specific machine-learning based approach using a Hamming distance based loss function outperforms the state-of-the-art classical approach. During further investigation, we discovered shortcomings in this approach, eventually leading to weaknesses in the entire system: The generated keys for the two legitimate partners either exhibit high entropy but fail to match (resulting in a high Hamming distance), or they match (yielding a low Hamming distance) but possess low entropy and remain constant. Both possible results render the keys useless for further use in (symmetric) cryptographic systems. We provide examples for our investigations and explanations regarding the weaknesses.