Disruptive approaches to handwriting and signature authentication for security-enhanced schemes
Rubén Tolosana Moranchel · Dialnet (Universidad de la Rioja) · 2019
Handwritten signature is one of the most socially accepted biometric traits as it has been used in financial and legal agreements for over a century. However, is signature biometric technology really adapted to current scenarios? With the massive deployment of mobile general purpose devices such as smartphones and tablets, new very interesting and user-friendly scenarios have appeared beyond the traditional office-like scenario considering high quality devices especifically designed for signature acquisition. In addition, despite the high technological evolution, and concretely, the success of deep learning techniques in combination with Graphics Processing Units (GPUs), the core of most of the state-of-the-art signature verification systems is still almost the same than 20 years ago. Why deep learning techniques do not outperform traditional systems as it happens in other fields? The last motivation for this Thesis is related to password-based systems. Traditionally, the two most prevalent user authentication approaches have been Personal Identification Numbers (PIN) and One-Time Passwords (OTP). However, and despite the high popularity and deployment of PIN- and OTP-based authentication systems in real scenarios, many studies have highlighted the weaknesses of these approaches as they are very easy to guess or steal (i.e., through shoulder-surfing and smudge attacks). Is it possible to increase the security of these traditional authentication systems at the same time that we provide a good experience to the users? As a way of finding the answers to these questions, this Thesis is mainly focused on the analysis of the new opportunities that bring up these novel scenarios and technologies and the challenges that must be tackled in order to achieve state-of-the-art results. This Dissertation comprises five different parts. Part I first concentrates on the problem statement and main contributions of the Thesis. The experimental chapters are then divided into three parts, Part II, Part III, and Part IV. Lastly, Part V concludes the Thesis. Part I first introduces the basics of biometrics, focusing on handwritten signature biometrics, which is the main topic of study in this Thesis, and the challenges and opportunities for it along an exhaustive overview of the state-of-the-art. Then, we concentrate on describing the most relevant features of existing on-line signature databases, making special emphasis on all the databases acquired during this Thesis. Finally, Part I concludes explaining first the specific details of the traditional on-line signature verification systems considered in the experimental parts of the Thesis, and then our novel end-to-end writer-independent RNN signature verification systems proposed in this Dissertation. The first experimental part (Part II of this Dissertation) starts analysing the system performance of traditional signature verification systems on emerging scenarios such as finger input, device interoperability and mixed writing-input. Due to the high system performance degradation of them, in this Thesis we propose a two-stage approach based on robust preprocessing and feature selection techniques. We then study the novel scenario where the number of stored samples or templates per user can grow very fast, making it possible to train more robust statistical user models, improving the performance of biometric systems, and in particular, reducing the template aging effect. The research carried out in this part aims to answer the following questions: How is the system performance affected on these novel scenarios? What approach should we consider to overcome these challenges? In the second experimental part (Part III of this Dissertation) we propose new ways to improve traditional signature verification systems. Concretely, we first evaluate the potential of including deep learning technology through a new architecture (Siamese) more adapted to the signature verification task. We then focus on the concept of complexity in signature and enhance the traditional systems through the selection of the most robust features for each signature complexity level. Finally, Part IV of this Dissertation evaluates the potential of incorporating handwriting biometric information to traditional authentication systems based on passwords, asking the user to draw each digit of the password on the touchscreen instead of typing them as usual. The research carried out in this Dissertation has led to novel contributions which include: i) analysis and adaptation of on-line signature verification systems to emerging scenarios such as finger input, device interoperability and mixed writing-input through robust preprocessing and feature selection techniques, ii) an exhaustive experimental analysis of template update strategies for three popular on-line signature verification approaches, extracting various practical findings related to the template aging effect in signature biometrics, and configuring time-adaptive improved versions of the considered baseline approaches overcoming to some extent the template aging, iii) exploring the potential of deep learning approaches for on-line signature verification. We have proposed a novel end-to-end writer-independent on-line signature verification system based on Recurrent Neural Networks with a Siamese architecture, which has outperformed other state-of-the-art systems, iv) improvement of traditional signature verification systems through the incorporation of the signature complexity concept, v) enhancement of traditional PIN and OTP authentication systems through the incorporation of handwriting biometric information as a second level of user authentication, vi) acquisition of new unprecedented handwriting and signature databases and release of them to the research community, and vii) part of the research presented in this Thesis has been deployed successfully in a pilot project in which on-line signature verification will be used massively in the Spanish banking sector.