Multitask GRU Learning for Channel Disturbance Detection, Cause Classification, and QBER Forecasting in Simulated Decoy-State BB84

Supervising a QKD link involves identifying variations in channel behaviour, while also considering noise arising from finite sampling. They can also be used to predict future values of QBER. A PyTorch gated recurrent unit (GRU) model is explored which includes a shared sequence encoder taking the recently reported count and QBER values. This GRU model includes three heads, one for detecting disturbances, another for determining the cause of the disturbance (multi-label classification), and the third for predicting QBER. 800 training sessions, 160 validation sessions and 320 test sessions were generated using the decoy-state BB84 simulator. All models contain 5 observed features and 3975 trainable parameters. The average recall per interval was 0.9622 across the three training seeds, with a false positive rate of 0.0081. The macro-F1 score for determining the cause was 0.9837. The prediction error for horizon 1 was 0.4762, for horizon 5 was 0.7055 and for horizon 10 was 1.0066. The model raised at least one alarm for all test events but not all disturbed intervals were detected. The model was tested to ensure the integrity of the guarded calibration, logit based classification loss and bounded predictions. However, slow drifting and transition between events proved challenging. These results show that the model can be used for joint supervised supervision in this simulator. However, it cannot be guaranteed to perform well on physical links, causes not seen during training will not be detected and does not guarantee security of the QKD.