Online Signature Verification Based on Generative Models

Area of Science:

• Biometrics and Pattern Recognition
• Machine Learning for Security Applications

Background:

• Generative models, particularly hidden Markov models (HMMs), are increasingly successful in online signature verification.
• Existing systems utilize user-specific HMMs (US-HMMs) and user-adapted universal background models (UA-UBMs), deriving scores from likelihood ratios and Viterbi path distances.

Purpose of the Study:

• To analyze the impact of feature set selection and HMM complexity on the performance of generative model-based signature verification systems.
• To investigate the role of dynamic information order, inclination angles, and pressure in feature sets.
• To explore fusion strategies for improved verification accuracy.

Main Methods:

• Experiments were conducted on the MCYT-100 database using skilled forgeries.
• Analysis focused on the influence of feature set dynamics (order, inclusion of angles/pressure) and HMM complexity.
• Verification scores were evaluated using likelihood ratios and Viterbi path distance.

Main Results:

• Viterbi path stability was observed across most feature sets and systems.
• US-HMM systems performed better with low-order dynamics in feature sets for likelihood evidence.
• UA-UBM systems showed superior results with likelihood ratios when low-order dynamics were excluded.
• Score-level fusion of US-HMM Viterbi path information and UA-UBM likelihood ratios enhanced verification performance.

Conclusions:

• Feature set composition and HMM complexity are critical factors influencing signature verification performance.
• Optimal feature dynamics depend on the specific HMM approach (US-HMM vs. UA-UBM).
• Fusion of complementary information from US-HMM and UA-UBM systems offers a promising path to higher accuracy in online signature verification.