Machine learning competition
We propose a competition to evaluate biometrical authentication based on smartphones’ inertial sensor data using machine learning techniques. To make the challenge more appealing, this problem will be presented as a crime scene investigation, where a smartphone belonging to a criminal was found as evidence and the registered sensor data needs to be compared to several suspects in custody.
A large train dataset of multiple people (suspects, including the criminal) using their smartphones naturally for long periods of time will be provided in advance, containing time series of accelerometer, gyroscope and magnetometer data at 100Hz. On the day of the competition, a test dataset will be provided, recorded under the same conditions as the train dataset but containing data of a single person (criminal).
The main purpose of the competition is to match test data to a single subject from the train dataset. The results will be based not only on the algorithms score given to the correct subject but also on the scores associated with the wrong people, to ensure that ambiguous results are not being generated.
While the results of this main problem will be presented in their own separate table, additional challenges will be unlocked as each team progresses. These challenges will ask the participants to detect additional traits or habits of each subject of the train dataset, such as being right or left-handed, which pocket is generally used to carry the smartphone or the respective age bracket and specific occurrences like sneezes or moments when pairs of subjects met each other. These challenges will count towards a separate score table with separate winners. Some of these challenges may require collecting additional data on-site by the participants.
This competition should be presented in a single category, single day format, with the main problem being the vehicle for research groups to showcase the efficacy of their biometric authentication algorithms, while the additional challenges can be seen as an optional robustness evaluation of each team.