Bio-crypto Systems Based on the Offline Signature Images


George Eskander, Ph.D candidate

Introduction

Bio-cryptography has been introduced to alleviate the key management problem of cryptography systems by using biometric traits to secure the private keys [1]. Among the different schemes, key-binding is generally the most reliable scheme. A challenging problem of such schemes is the fuzziness of the biometric signals that results from the intra-personal variability and the interpersonal similarity that may cause rejection of authorized users and acceptance of unauthorized users respectively. Among various key-binding schemes, the fuzzy vault (FV) scheme may deal efficiently with the fuzzy nature of biometric signals [2]. This research proposal aims to design a reliable bio-cryptography system based on the offline handwritten signature images and the FV construction that may facilitate information security applications and secure handwritten signature verification systems.

Problem statement

A FV system operates like a biometric signal classifier. The same challenges to design a biometric recognition system, namely, recognition quality, security, and resource management apply to FV schemes. The quality of recognition depends on the feature vector used to encode the FV, and on the parameters of the FV system itself. A feature vector should be selected from the biometric trait template; this vector should discriminate between the different system users and also between the genuine and impostor persons, and should be relatively invariant for the same person. For the offline handwritten signatures, variability between samples make it hard to extract stable features. Also, if some stable features exist, it may be similar for different users. Selecting good features is a trade-off between system accuracy and robustness; selecting very discriminating features guarantees low False Acceptance Rate (FAR), and hence increases the system accuracy. But, it may increase the False Reject Rate (FRR), and hence decrease the system robustness. On the other hand, selecting very stable features guarantees the system robustness but may decrease its accuracy.

Achieving a very informative feature vector is not the only factor controls the FV recognition quality as FV parameters like the degree of the encoding polynomial, encoding size, and the quantity of added noise (chaff) are affecting the recognition performance. It is a challenging task to discover the intrinsic properties of features and chaff points along with the FV parameters that lead to the optimal trade-off between the system accuracy, and robustness; taking in consideration the other performance measures like system security, speed, and memory requirements.

Methodology

The proposed research solves such a complicated design problem through two stages. The first stage aims to design a single FV system that utilizes a subset of the most informative features. The second stage aims to design an ensemble of FV systems; each utilizes a different subset of features and/or has different design parameters, which cooperate to take the final recognition decision based on a decision fusion mechanism.

In the first stage, the interpersonal similarity and intra-personal variability problems are tackled from the feature selection perspective by applying the similarity learning approach [3]. This approach relies on representing the features in the dissimilarity space by using the absolute distance between two features as a new representative feature. The dissimilarity representation of a feature reflects its stability and accuracy. Therefore, features can be ranked according to these two quality indexes. After representing the features in the dissimilarity space, specific feature selection algorithms should be applied. For instance, boosting feature selection algorithms, e.g. AdaBoost [4], are able to rank the features by giving a weight to each feature based on its classification ability.

The second stage of the proposed system aims to design an ensemble of FV systems by fusing the FVs individual decisions. Instead of trying to encode a single very accurate FV system using the optimal set of features and the optimal settings of FV parameters, multiple FVs will be encoded each utilizes different set of features and/or has different design parameters. In the decoding time, all FVs will be decoded and a decision fusion rule will be applied in the ROC space to judge the different outputs [5].

Applications

Research results will be validated by integrating the proposed framework in the development of a project supported by BancTec Inc. The project aims to develop an automated system for securing a massive amount of financial documents by applying a combination of intelligent watermarking and bio-cryptography technologies; as it is a compromising approach to enhance the information security systems by combine biometrics, cryptography and data hiding, i.e., watermarking, methodologies [6]. The proposed system will be used to achieve the confidentiality, integrity, and authenticity of the digitized handwritten documents, e.g., bank cheques, by utilizing its embedded signatures as biometric keys by which classical keys of different security frameworks, e.g. encryption, watermarking, etc, can be managed.

Publication

George Eskander, Eric Granger, and Robert Sabourin, Signature Based Fuzzy Vaults with Boosted Feature Selection, Submitted to IEEE Workshop on Computational Intelligence and Identity Management (CIBIM 2011).

References

[1] U. Uludag, S. Pankanti, S. Prabhakar, A.K. Jain. Biometric Cryptosystems: Issues and Challenges. Proceedings of IEEE, vol.92, pp.948-960, 2004.
[2] A. Juels, M. Sudan. A Fuzzy Vault Scheme. Proceedings of IEEE int. Symp. Inf. Theory, Lausanne, Switzerland, pp.408, 2002.
[3] S. Cha. Use of Distance Measures in Handwriting Analysis. PhD thesis, State University of New York at Buffalo, 2001.
[4] Schapire et al., The boosting approach to machine learning: An overview. Workshop on Nonlinear Estimation and Classification, 2002.
[5] W. Khreich, Eric Granger, Ali Miri and Robert Sabourin. Iterative Boolean Combination of Classifiers in the ROC Space: An Application to Anomaly Detection with HMMs. Pattern Recognition, vol.43 , Issue. 8, pp.2732-2752, August 2010.]
[6] J. Dong, and T. Tan. Security Enhancement of Biometrics, Cryptography and Data Hiding by Their Combinations. 5th International Conference on Visual Information Engineering , 2008.