How on earth do you even get training data for this? Recorded phone sensor outputs that are known for certain to be the result of validated, confirmed theft events can't possibly be that common. Are they paying people in Bangledesh a few bucks to be randomly assigned to group that either get robbed or tripped in the hope they throw the phone and labeling sensor data accordingly. When this type of motion recognition was first developed, they had labs and recorded people walking, doing jumping jacks, sitting and then standing, whatever, and labeled the patterns appropriately because they knew what was happening because it was happening in a lab.