Return difference in uncertainties for model-uncertainty detectors

[ojcobb]

For the model-uncertainty detectors returning the difference in average uncertainty on the ref set vs the test set can give a good indication of whether the drift is likely to be malicious or not. The idea is that an increase in uncertainty implies that the model performance is likely decreasing, whilst a decrease in uncertainty indicates it might actually be increasing. This wouldn't be perfect however as models can be overconfident on OOD instances and so it's possible that average uncertainty might decrease under malicious drift.

The implementation is non-trivial as currently the model-uncertainty detectors wrap around KS or Chi^2 detectors which compute uncertainties as an intermediary step and don't return them.