In response to detecting the trigger event, the system starts data collection (step 504) and sometime later stops data collection (step 506). (Note that in some embodiments, data collection does not necessarily stop before the data is processed. It can be processed in a streaming manner (or in chunks) while the data is still being collected.) Then, the system performs signal-processing operations on the data to produce a feature vector (step 508). The system then uses a model generated using machine-learning techniques (such as a neural network) to process the feature vector (step 510). Note that the system may generate a number of models for each user, wherein each model is associated with a specific behavior, such as walking, standing up or sitting down. The system can also generate a “universal background model,” which ideally includes characteristics for the entire human race, and can determine how the user fits into the universal background model. In particular, the system can use the universal background model to identify “similar people” who exhibit characteristics, which are closest to the user, and can construct synthetic training data to train the model to discriminate between the user and the similar people.

Finally, the system generates a security score for the user along with an associated confidence value (step 512). Note that while determining the security score, the system can generate a “generalized product of experts.” For example, the system can include a collection of experts, such as an expert that determines whether sensor data matches a user's gait, and another expert that determines a user's location. The system can then use inputs from all of these experts to generate the security score.