Polar codes have been adopted as the coding scheme for control channels in the 5G communication system, where the blind detection is required. In this letter, to mitigate the false-alarm rate (FAR) in the polar code blind detection, we propose a more efficient architecture which combines the conventional list decoding and a binary classifier. After the list decoding, the CRC-passing result will be further checked by the classifier to determine whether the received signal is the intended control information or not. The classifier works by inspecting the squared Euclidean distance ratio (SEDR) between the received signal and the hypotheses, and the decision threshold used in the classifier is automatically learnt by a neural network (NN). The presented results show that very short CRC sequences can be enough to reach the target FAR, and this CRC overhead reduction finally contributes to the performance improvement of polar codes.