Machine Learning Assisted Wiretapping

It is well known, that wiretap codes can be used to protect against a potential eavesdropper in a communication scenario. Asymptotically, they can achieve both vanishing decoding error probability at a legitimate receiver and vanishing leaked information to an eavesdropper. However, under finite blocklength, this does not hold and there is a tradeoff between different code parameters. In this work, it is shown, how machine learning algorithms can be utilized by an eavesdropper to decode finite-blocklength polar wiretap codes successfully. Neural networks and support vector machines are implemented as channel decoders and compared to a reference polar decoder. Simulation results show that the support vector machine decoders can outperform the reference decoder with respect to the bit error ratio.