Finger-Drawn Signature Verification on Touch Devices Using Statistical Anomaly Detectors

The use of behavioral biometrics in user authentication has recently moved to new security application areas, one of which is verifying finger-drawn signatures and PIN codes. This paper investigates the design of anomaly detectors and feature sets for graphic signature authentication on touch devices. The work involved a selection of raw data feature sets that are extracted from modern mobile devices, such as finger area, pressure, velocity, acceleration, gyroscope, timestamp and position coordinates. A set of computed authentication features are formulated, derived from the raw features. The proposed anomaly detector is based on the outlier method, using three versions of the Z-Score distance metric. The proposed feature sets and anomaly detectors are implemented as a data collection and dynamic authentication system on an Android tablet. Experimental work resulted in collecting a signature dataset that included genuine and forged signatures. The dataset was analyzed using the Equal-Error-Rate (EER) metric. The results for random forgery and skilled forgery showed that the Z-Score anomaly detector with 3.5 standard deviations distance from the mean produced the lowest error rates. The skilled forgery error rates were close to random forgery error rates, indicating that behavioral biometrics are the key factors in detecting forgeries, regardless of pre-knowledge of the signature's shape.