Study of long-term quality of online signature verification systems

Real handwriting authentication systems need a robust writer identification over a long time period. The paper analyzes signature sessions of the ATV-Signature Long Term Database (ATV-SLT DB). The database contains 6 sessions generated by 27 users over 15 month. The quality change of the verification results over a period of 15 month is examined. 64static and dynamic biometric features from the ATV-SLT DB sessions are extracted and 3 different classifiers are used. For the impostor test a 7th session is added, the impostor session, with 6 signatures for each user. The best result of 99.17% success rate for a correct classification is reached with the k-Nearest Neighbor classifier. The best result of 2.47% false accepted rate is reached with Naïve Bayes classifier.