论文部分内容阅读
Abstract: Multiuser online system is useful, but the administrator must be nervous at security problem. To solve this problem, the authors propose applying signature verification to multiuser online system. At the authors’ research, they attempt adding signature verification function based on DP (Dynamic Programming) matching to existing multiuser online kanji learning system. In this paper, the authors propose the construction of the advance system and methods of signature verification, and evaluate performance of those signature verification methods that difference is combination of using features. From signature verification’s experimental results, the authors adopted to use writing velocity and writing speed differential as using feature to verify the writer for the system. By using signature database which is construct with 20 genuine signatures and 20 forged signatures with 40 writers and written mostly by English or Chinese literal, experimental results of signature verification records 12.71% as maximum EER (Equal Error Rate), 6.00% as minimum EER, and 8.22% as average EER. From mentioned above, the authors realized to advance the reliability and usefulness of the multiuser online kanji learning system.
Key words: Signature verification, character learning method, dynamic programming.
1. Introduction
To write Japanese, kanji is one of the most popular letters which is as same as hiragana and katakana, therefore to live in Japan, acquirement of reading and writing kanji must be required. These days, Japanese Ministry of Education announced 1,006 kanji as kyoiku kanji which must be acquired in 6 years during elementary school, 2,136 kanji as joyo kanji which must be acquired if living in Japan. Most Japanese will study huge amount of kanji, as like mentioned above, through lesson in educational facilities, such as elementary school and junior high school. However, the lesson time in each educational facility is limited, we could not say necessary that all students can acquire all kanji in the lesson. On the other hand, many foreign people are living in Japan because of the globalization of economy, so the acquirement of kanji is also must need for them. Therefore, many people are hoping to use the tool which could learn kanji effectively in short term and not limit when or where.
Nowadays, many kanji learning system which is based on computer has researched and developed by many researchers as a kanji learning tool. System developed in Ref. [1] focused user as foreign people or elementary school student. Using system developed in Ref. [2], teacher could customize dictation’s evaluation point and reference kanji input data. System developed in Ref. [3] compliant to network communication.
We have also researched kanji learning system in past long time. At first, the basis of the kanji learning system has been constructed [4]. Among the past research, the system has been improved to multiuser online system [5]. By this improvement, our kanji learning system becomes more useful for using in educational facility.
As like mentioned above, existing kanji learning system has very high performance. However, the system has including risk that it is easy for someone else to pass the login security if the password has been stolen. This means when user use existing system it will be possible to cause improper result at scoring kanji learning or getting the data of human kanji learning progress which could use at other research, not only influence on personal kanji learning. To avoid such risks and decrease lawbreaker’s login, we will propose to implement signature verification as self verification at login function to the existing kanji learning system. Signature which will use in signature verification is one of the human’s biometric features that are very personal and hard to copy, unlike password that could use by everyone. Changing self verification method will realize another advantage. If we change self verification method to signature verification at the system, the number of input device will be decrease. This means user has not need to use keyboard, only pen device to input login key so that we could expect advancement of the system’s usefulness.
This paper is composed with six sections. In the second section, we will explain about construction and summary of each application of existing multiuser online kanji learning system. In the third section, we will explain about attention points that will occur when the system’s self verification method is changed to signature verification. In the fourth section, we will state about signature verification method which is implement to the system. In the fifth section, we will discuss about experimental result of signature verification which is mentioned in section 4. In the last section, we will conclude the research.
2. Kanji Learning System
Existing kanji learning system in our laboratory has been researched and improved with many people. The basis of the system has been constructed by Ref. [4]. The origin system has not only includes kanji correction function such as shape correction and writing order correction, it has also includes kanji writing animation function which will help users to understand correct shape and writing order of the learning kanji. Improvement of the system has been continued little by little, and the latest system realized every time and everywhere kanji learning [5]. The system could mainly separate into two parts, client side and server side. Client side has relationship between client users such as students in educational facility. The user will use student application which base is origin kanji learning system to learn kanji. On the other hand, server side has relationship between master users such as teachers in educational facility. The user will use teacher application to supervise clients’ kanji learning. By the realization of network connection between client side and server side, latest kanji learning system becomes online system and it has materialized more effective kanji learning. The latest existing system has also includes scoring function that will automatically score the user’s written kanji. Therefore, the use of our kanji learning system in educational facility becomes more practical, that is one of the goals for our research.
3. Login Security
As like mentioned in section 2, existing kanji learning system realize effective kanji leaning by many useful functions and free kanji learning, not limited with time or place, by implementing network connection. However, the system adopted to use password which is non-personal as self verification when client request connection to server. This will cause dishonest login because if the password is known by outsider, they could easily copy and use the password to pass the self verification. Impropriety result cause from dishonest login is not only influence on personal kanji learning. For example, we could expect improper result at scoring kanji learning and getting the data of kanji learning progress. Scoring kanji learning will be indispensable if the system is used in educational facilities lesson and getting kanji learning progress data will be useful for analyzing human’s kanji learning progress.
To not realize such risks and make closely impossible to success the dishonest login, we will propose to implement signature verification as self verification at login function to the existing kanji learning system. Signature verification will verify the client is self or not by using signature which is one of the human’s biometric features. Unlike password which could easily copy and use by everyone, biometric feature is different with one by one and difficult to copy by others. Therefore, it will be closely impossible for someone else to pass the system’s login security even if he/she has known the reasonable signature. Changing self verification method will realize another advantage, not only mentioned before. If we change self verification method to signature verification at the system, user has not need to use keyboard, just only using pen device to input login key, so the user could use the system with only controlling pen device. Considering the fact that elementary schoolchild who is one of the main targets of the system could not control the keyboard perfectly [6], it could be consider as advantage that the user will be possible to learn kanji just only with pen device.
To change existing multiuser online kanji learning system’s self verification method to signature verification, we revise the following points at the system:
(1) Change using personal information for login:
Exist: name, class number, student identity number (ID), birthday, password;
Proposed: class number, student ID, signature.
(2) Register some personal signatures as training data for signature verification to the server;
(3) Add signature verification function to the server.
Fig. 1 shows the construction of the new online kanji learning system. Moreover, the part which is printed with green is fixed and added from the existing system because self verification method becomes to signature verification.
Steps to Login:
Because method for client login at kanji learning system is changed, process for the login will also changes. Following shows the steps of new login process. Fig. 2 shows the passage of data and process at the new online kanji learning system.
Step 1. Client: Write signature in designate area at student application.
Step 2. Client: Input personal information as like as existing system but without name, birthday and password.
Step 3. Client: Send personal information and signature data to the server.
Step 4. Server: Compare received signature data with reference signature data by signature verification.
Step 5. Server: If signature verification’s result says the client is him/herself, change the state as attendance and send message to the client that login was successful. On the other hand, the result says the client is dishonest send message to the client that login was failed.
4. Signature Verification
In this section, we will explain about signature verification and the method we adopted to the system. Signature verification is one of biometric verification and it will use human’s signature to verify the writer is him/herself or outsider. Unlike password, biometric is very personal and useful because it is feature of someone’s looking or behavior. Furthermore, biometric is hard to copy so it is good feature for verification.
Signature verification can be classified into two big types. One is offline signature verification and another is online signature verification. Offline signature verification is signature verification which target is the signature already written. However, it will use only shape information of the signature that is easily to copy so possibility of false acceptation by outsider will increase. On the other hand, online signature verification will use transition of human’s writing. Transition of writing velocity, pressure and so on is invisibility data so it is hard to copy for outsider. As mentioned above, online signature verification realizes higher security performance and it could use various features so we will adopt it to the multiuser online kanji learning system.
Nowadays, many methods had proposed for online signature verification. For example, there is method which will attend to pen inclination [7]. Another method will use Fourier descriptors for signature verification [8]. Furthermore, there is method that will use cancelable templates to increase security [9].
To verify input signature is written by himself/herself or not, we need to make correspondence between two data strings such as input and reference to compute similarity of the input. The method Dynamic Programming (DP) matching is one of popular method in pattern matching. In this research, we use DP matching to find matching pattern between two data strings.
4.1 DP Matching
DP matching is one of the pattern matching methods that will make correspondence between two data strings. Using DP matching, we could know the similarity between two data strings as the distance. Unlike linear matching, DP matching will realize more suitable matching result between plural data strings. Fig. 3 shows the example of matching result between two data strings. As you can see, the distance which is measurement of difference between two data strings by DP matching is much smaller than distance by linear matching.
General idea of DP matching is to make shortest path between two opposite points at grid map which example is shown in Fig. 4. At this time, ), ( jigis minimum distance between start point to grid point (i, j) and it will be decided by slope constraint. Therefore, distance between two data strings A and B that each length is I and J is ).,(),(JIgBAd= In this research, we established following slope constraint for the system.
4.2 Verification Method
As mentioned above, we succeed to get distance between two data strings by DP matching. However, some problem will occur if we use the distance as similarity between the data with unprocessed. It is okay if the signature verification system is exclusive for one signature pattern because if the distance will be smaller it means two data strings is much similar, but the system is supposed to use many signature pattern.
There is a tendency that the distance which calculated by DP matching will be smaller if length of data strings is shorter so we need to process the distance to get the similarity between two signatures.
To solve this problem, we will calculate signature’s similarity by using distance between training data [10]. Training data is some reference signature data which is written by genuine user and registered to server beforehand. The steps for self verification will be process as following. First of all, the system will calculate the distance between each training data, the number of it is NREF, and calculate some value from the result. Secondly, calculate D which is the similarity between testing signature data and genuine signature data by using the value which calculated in step before. Finally, verify the test signature is written by him/herself or not by magnitude relationship between D and threshold θ.
At this system, test signature will be judged as self signature when θ≤ D and it will be judged as outsider’s signature when other case. In this paper, we tested following factors to calculate similarity between training and test data, M1: standard score from training data, M2: average distance between training data, M3: minimum distance between training data, M4: maximum distance between training data, standard score of training data. Following is the expression to calculate similarity of input signature data. By this time,μ is the average distance between training data, σ is the standard division between training data, τ is the value of factor for the method.
For M1)(3)
The self verification system is only uses xy coordinates among the data which is possible to acquire from pen device. This is because existing kanji learning system also uses only coordinate information and we want to reduce the amount of data which will send from client side to server side throw network. From acquired coordinate information, the system could get several data such as writing velocity v, writing speed s and writing speed differential s’ as features which could use for signature verification. Moreover, to calculate such feature data, following expressions will be use.])
5. Experimental Result
In this section, we will indicate and discuss about experimental result of the signature verification which method is presented in section 4. For the experiment, we use SVC2004 Task2 [11] database to test the signature verification. SVC2004 Task2 is one of the signature databases and it is constructed by 20 genuine signatures and 20 forged signatures with 40 writers. The database’s signature data is mostly written by
English or Chinese. Because one of the our multiuser online kanji learning system’s main target is foreign people and Japanese kanji is similar to Chinese literal, the database is suitable for this experiment.
At this experiment, signature verification system will use five randomly selected genuine signatures as training data, remainder 15 genuine signatures and 20 forged signatures as test data, for each signature pattern. Therefore, the system will verify 600 genuine signatures and 800 forged signatures for one experiment. In this paper, we will write the result that is acquired from eight experiments those training pattern will change with each experiment.
Fig. 5 shows signature verification result with single feature. FRR is stands for false reject rate and it shows percentage of verification result which reject genuine signature by mistake. FAR is stands for false accept rate and it shows percentage of verification result which accept forged signature by mistake. EER is stands for equal error rate and it shows percentage of verification result when FRR and FAR is balance. We could see that method which will calculate signature’s similarity by using average distance between training data has record best verification result for each feature. As a result, we say that using method which will calculate signature’s similarity by using average distance between training data is most suitable for signature verification, so that we will continue signature verification experiments as like it.
To get better signature verification result, we test signature verification which will verify signature with similarity that is combination of single similarity. Complex similarity will calculated as following expression. In this time, we will express Dv for single similarity by writing velocity, Ds for single similarity by writing speed, Ds’ for single similarity by writing speed differential, DC1 for complex similarity by writing velocity and speed, DC2 for complex similarity by writing velocity and speed differential, DC3 for complex similarity by writing velocity, speed and speed differential.
(5)
The signature verification results with complex similarity are shown in Fig. 6-8. Fig. 6 shows signature verification result with complex similarity by writing velocity and speed. The best EER result records 8.33% when α = 0.7 and β = 0.3. Fig. 7 shows signature verification result with complex similarity by writing velocity and speed differential. The best EER result records 8.22% when α = 0.45 and β = 0.55. Fig. 8 shows signature verification result with complex similarity by writing velocity, speed and speed differential. The best EER result records 8.18% when α= 0.4, β = 0.1, γ = 0.5 or α = 0.5, β = 0.2, γ = 0.3. Furthermore, experiments when α = 0.6, β = 0.2, γ = 0.2 and α = 0.7, β = 0.2, γ = 0.1 records closely best EER result.
The best signature verification result was calculated as mentioned before. To use the system practical, we must use the threshold which has been fixed. Therefore, we calculate the verification rate which is using common threshold for each eight experiment. In this case, we set the fixed threshold as average of suitable threshold for each experiment. Fig. 9 shows result of signature verification. Similarity of signature will be calculated as following expression.
(6)
As a result, the signature verification method which records best performance was method that uses complex similarity by writing velocity and speed differential. The signature verification method records 10.29% as average FRR, 9.72% as average FAR and 9.96% as total error rate.
6. Conclusions
In this paper, we proposed to change self verification method as signature verification from verification method which use password at existing multiuser online kanji learning system. By adopting signature verification, we succeed to improve reliability and usability of the system. From experimental result, we adopted signature verification to use writing velocity and speed differential as feature. We adopted it because there was no big difference between the verification rate which records lowest EER such as 8.18% and it does not need to calculate the similarity by writing speed so the verification time will be decrease. Average EER of the adopting method was 8.22% when we establish threshold most suitable. On the other hand, average of FRR records 10.29%, FAR records 9.72% and total error rate records 9.96% when we establish threshold as constant, to simulate if the signature verification function is implement in actual running system. From mentioned above, we could say new multiuser online kanji learning system become more practical when we compare with existing system.
As future works, we could consider following improvements to realize the widely use of the system. First, reinforcement of the communication tools between server and client to realize easy and deep understanding at kanji learning. Second, addition of the function that could statistically compute the tendency of each user’s weak points at writing kanji, radical, and so on. The addition of this function will realize more effective kanji
learning. Finally we perform the improvement of signature verification method. This will realize construction of more reliable multiuser online system.
References
[1] R. Tatsuoka, M. Yoshimura, Development of a kanji learning system for foreign students or elementary students, IEICE Trans. & System, ET96-36, June 1996.
[2] N. Takesue, K. Mochida, A. Kitadai, M. Nakagawa, A handwriting-based kanji learning system enabling teachers to designate evaluation points, IPSJ SIG Technical Reports (15) (2005) 15-22.
[3] T. Iwasaki, A. Kitadai, J. Tokuno, M. Nakagawa, A web based training system to learn kanji, IPSJ SIG Technical Reports 16 (2006) 17-24.
[4] A. Takeda, J. Shin, Character Learning System Using Inter-stroke Information, in: Proceedings of the 8th International Conference on Knowledge-Based Intelligent Information, Wellington, New Zealand, Sep. 2004, pp. 165-174.
[5] Y. Shimizu, J. Shin, User friendly kanji learning system and its evaluation, Master’s Thesis, University of Aizu, Mar. 2010.
[6] J. Takahashi, Status of primary schoolchild’s keyboard skill through development of keyboard input study web site (in Japanese), Univ. of Toyama Information Technology Center Public and Information, 2004, Vol. 1, pp. 16-19.
[7] S. Nakajima, T. Hamamoto, S. Hangai, On-line signature verification using pen inclination, IEICE (1998) 15-22.
[8] B. Yanikoglu, A. Kholmatov, Online signature verification using fourier descriptors, EURASIP Journal on Advances in Signal Processing 2009 (2009).
[9] E. Maiorana, P. Campisi, A. Neri, Template protection for dynamic time warping based biometric signature authentication, in: 16th International Conference on Digital Signal Processing (DSP 2009), Santorini, Greece.
[10] W.-D. Chang, J. Shin, DPW approach for random forgery problem in online handwritten signature verification, in: 4th International Conference on Networked Computing and Advanced Information Management, Korea Republic, 2008.
[11] D.-Y. Yeung, H. Chang, Y. Xiong, S. George, R. Kashi, T. Matsumoto, G. Rigoll, SVC2004: First international signature verification competition, ICBA, July 2004.
Key words: Signature verification, character learning method, dynamic programming.
1. Introduction
To write Japanese, kanji is one of the most popular letters which is as same as hiragana and katakana, therefore to live in Japan, acquirement of reading and writing kanji must be required. These days, Japanese Ministry of Education announced 1,006 kanji as kyoiku kanji which must be acquired in 6 years during elementary school, 2,136 kanji as joyo kanji which must be acquired if living in Japan. Most Japanese will study huge amount of kanji, as like mentioned above, through lesson in educational facilities, such as elementary school and junior high school. However, the lesson time in each educational facility is limited, we could not say necessary that all students can acquire all kanji in the lesson. On the other hand, many foreign people are living in Japan because of the globalization of economy, so the acquirement of kanji is also must need for them. Therefore, many people are hoping to use the tool which could learn kanji effectively in short term and not limit when or where.
Nowadays, many kanji learning system which is based on computer has researched and developed by many researchers as a kanji learning tool. System developed in Ref. [1] focused user as foreign people or elementary school student. Using system developed in Ref. [2], teacher could customize dictation’s evaluation point and reference kanji input data. System developed in Ref. [3] compliant to network communication.
We have also researched kanji learning system in past long time. At first, the basis of the kanji learning system has been constructed [4]. Among the past research, the system has been improved to multiuser online system [5]. By this improvement, our kanji learning system becomes more useful for using in educational facility.
As like mentioned above, existing kanji learning system has very high performance. However, the system has including risk that it is easy for someone else to pass the login security if the password has been stolen. This means when user use existing system it will be possible to cause improper result at scoring kanji learning or getting the data of human kanji learning progress which could use at other research, not only influence on personal kanji learning. To avoid such risks and decrease lawbreaker’s login, we will propose to implement signature verification as self verification at login function to the existing kanji learning system. Signature which will use in signature verification is one of the human’s biometric features that are very personal and hard to copy, unlike password that could use by everyone. Changing self verification method will realize another advantage. If we change self verification method to signature verification at the system, the number of input device will be decrease. This means user has not need to use keyboard, only pen device to input login key so that we could expect advancement of the system’s usefulness.
This paper is composed with six sections. In the second section, we will explain about construction and summary of each application of existing multiuser online kanji learning system. In the third section, we will explain about attention points that will occur when the system’s self verification method is changed to signature verification. In the fourth section, we will state about signature verification method which is implement to the system. In the fifth section, we will discuss about experimental result of signature verification which is mentioned in section 4. In the last section, we will conclude the research.
2. Kanji Learning System
Existing kanji learning system in our laboratory has been researched and improved with many people. The basis of the system has been constructed by Ref. [4]. The origin system has not only includes kanji correction function such as shape correction and writing order correction, it has also includes kanji writing animation function which will help users to understand correct shape and writing order of the learning kanji. Improvement of the system has been continued little by little, and the latest system realized every time and everywhere kanji learning [5]. The system could mainly separate into two parts, client side and server side. Client side has relationship between client users such as students in educational facility. The user will use student application which base is origin kanji learning system to learn kanji. On the other hand, server side has relationship between master users such as teachers in educational facility. The user will use teacher application to supervise clients’ kanji learning. By the realization of network connection between client side and server side, latest kanji learning system becomes online system and it has materialized more effective kanji learning. The latest existing system has also includes scoring function that will automatically score the user’s written kanji. Therefore, the use of our kanji learning system in educational facility becomes more practical, that is one of the goals for our research.
3. Login Security
As like mentioned in section 2, existing kanji learning system realize effective kanji leaning by many useful functions and free kanji learning, not limited with time or place, by implementing network connection. However, the system adopted to use password which is non-personal as self verification when client request connection to server. This will cause dishonest login because if the password is known by outsider, they could easily copy and use the password to pass the self verification. Impropriety result cause from dishonest login is not only influence on personal kanji learning. For example, we could expect improper result at scoring kanji learning and getting the data of kanji learning progress. Scoring kanji learning will be indispensable if the system is used in educational facilities lesson and getting kanji learning progress data will be useful for analyzing human’s kanji learning progress.
To not realize such risks and make closely impossible to success the dishonest login, we will propose to implement signature verification as self verification at login function to the existing kanji learning system. Signature verification will verify the client is self or not by using signature which is one of the human’s biometric features. Unlike password which could easily copy and use by everyone, biometric feature is different with one by one and difficult to copy by others. Therefore, it will be closely impossible for someone else to pass the system’s login security even if he/she has known the reasonable signature. Changing self verification method will realize another advantage, not only mentioned before. If we change self verification method to signature verification at the system, user has not need to use keyboard, just only using pen device to input login key, so the user could use the system with only controlling pen device. Considering the fact that elementary schoolchild who is one of the main targets of the system could not control the keyboard perfectly [6], it could be consider as advantage that the user will be possible to learn kanji just only with pen device.
To change existing multiuser online kanji learning system’s self verification method to signature verification, we revise the following points at the system:
(1) Change using personal information for login:
Exist: name, class number, student identity number (ID), birthday, password;
Proposed: class number, student ID, signature.
(2) Register some personal signatures as training data for signature verification to the server;
(3) Add signature verification function to the server.
Fig. 1 shows the construction of the new online kanji learning system. Moreover, the part which is printed with green is fixed and added from the existing system because self verification method becomes to signature verification.
Steps to Login:
Because method for client login at kanji learning system is changed, process for the login will also changes. Following shows the steps of new login process. Fig. 2 shows the passage of data and process at the new online kanji learning system.
Step 1. Client: Write signature in designate area at student application.
Step 2. Client: Input personal information as like as existing system but without name, birthday and password.
Step 3. Client: Send personal information and signature data to the server.
Step 4. Server: Compare received signature data with reference signature data by signature verification.
Step 5. Server: If signature verification’s result says the client is him/herself, change the state as attendance and send message to the client that login was successful. On the other hand, the result says the client is dishonest send message to the client that login was failed.
4. Signature Verification
In this section, we will explain about signature verification and the method we adopted to the system. Signature verification is one of biometric verification and it will use human’s signature to verify the writer is him/herself or outsider. Unlike password, biometric is very personal and useful because it is feature of someone’s looking or behavior. Furthermore, biometric is hard to copy so it is good feature for verification.
Signature verification can be classified into two big types. One is offline signature verification and another is online signature verification. Offline signature verification is signature verification which target is the signature already written. However, it will use only shape information of the signature that is easily to copy so possibility of false acceptation by outsider will increase. On the other hand, online signature verification will use transition of human’s writing. Transition of writing velocity, pressure and so on is invisibility data so it is hard to copy for outsider. As mentioned above, online signature verification realizes higher security performance and it could use various features so we will adopt it to the multiuser online kanji learning system.
Nowadays, many methods had proposed for online signature verification. For example, there is method which will attend to pen inclination [7]. Another method will use Fourier descriptors for signature verification [8]. Furthermore, there is method that will use cancelable templates to increase security [9].
To verify input signature is written by himself/herself or not, we need to make correspondence between two data strings such as input and reference to compute similarity of the input. The method Dynamic Programming (DP) matching is one of popular method in pattern matching. In this research, we use DP matching to find matching pattern between two data strings.
4.1 DP Matching
DP matching is one of the pattern matching methods that will make correspondence between two data strings. Using DP matching, we could know the similarity between two data strings as the distance. Unlike linear matching, DP matching will realize more suitable matching result between plural data strings. Fig. 3 shows the example of matching result between two data strings. As you can see, the distance which is measurement of difference between two data strings by DP matching is much smaller than distance by linear matching.
General idea of DP matching is to make shortest path between two opposite points at grid map which example is shown in Fig. 4. At this time, ), ( jigis minimum distance between start point to grid point (i, j) and it will be decided by slope constraint. Therefore, distance between two data strings A and B that each length is I and J is ).,(),(JIgBAd= In this research, we established following slope constraint for the system.
4.2 Verification Method
As mentioned above, we succeed to get distance between two data strings by DP matching. However, some problem will occur if we use the distance as similarity between the data with unprocessed. It is okay if the signature verification system is exclusive for one signature pattern because if the distance will be smaller it means two data strings is much similar, but the system is supposed to use many signature pattern.
There is a tendency that the distance which calculated by DP matching will be smaller if length of data strings is shorter so we need to process the distance to get the similarity between two signatures.
To solve this problem, we will calculate signature’s similarity by using distance between training data [10]. Training data is some reference signature data which is written by genuine user and registered to server beforehand. The steps for self verification will be process as following. First of all, the system will calculate the distance between each training data, the number of it is NREF, and calculate some value from the result. Secondly, calculate D which is the similarity between testing signature data and genuine signature data by using the value which calculated in step before. Finally, verify the test signature is written by him/herself or not by magnitude relationship between D and threshold θ.
At this system, test signature will be judged as self signature when θ≤ D and it will be judged as outsider’s signature when other case. In this paper, we tested following factors to calculate similarity between training and test data, M1: standard score from training data, M2: average distance between training data, M3: minimum distance between training data, M4: maximum distance between training data, standard score of training data. Following is the expression to calculate similarity of input signature data. By this time,μ is the average distance between training data, σ is the standard division between training data, τ is the value of factor for the method.
For M1)(3)
The self verification system is only uses xy coordinates among the data which is possible to acquire from pen device. This is because existing kanji learning system also uses only coordinate information and we want to reduce the amount of data which will send from client side to server side throw network. From acquired coordinate information, the system could get several data such as writing velocity v, writing speed s and writing speed differential s’ as features which could use for signature verification. Moreover, to calculate such feature data, following expressions will be use.])
5. Experimental Result
In this section, we will indicate and discuss about experimental result of the signature verification which method is presented in section 4. For the experiment, we use SVC2004 Task2 [11] database to test the signature verification. SVC2004 Task2 is one of the signature databases and it is constructed by 20 genuine signatures and 20 forged signatures with 40 writers. The database’s signature data is mostly written by
English or Chinese. Because one of the our multiuser online kanji learning system’s main target is foreign people and Japanese kanji is similar to Chinese literal, the database is suitable for this experiment.
At this experiment, signature verification system will use five randomly selected genuine signatures as training data, remainder 15 genuine signatures and 20 forged signatures as test data, for each signature pattern. Therefore, the system will verify 600 genuine signatures and 800 forged signatures for one experiment. In this paper, we will write the result that is acquired from eight experiments those training pattern will change with each experiment.
Fig. 5 shows signature verification result with single feature. FRR is stands for false reject rate and it shows percentage of verification result which reject genuine signature by mistake. FAR is stands for false accept rate and it shows percentage of verification result which accept forged signature by mistake. EER is stands for equal error rate and it shows percentage of verification result when FRR and FAR is balance. We could see that method which will calculate signature’s similarity by using average distance between training data has record best verification result for each feature. As a result, we say that using method which will calculate signature’s similarity by using average distance between training data is most suitable for signature verification, so that we will continue signature verification experiments as like it.
To get better signature verification result, we test signature verification which will verify signature with similarity that is combination of single similarity. Complex similarity will calculated as following expression. In this time, we will express Dv for single similarity by writing velocity, Ds for single similarity by writing speed, Ds’ for single similarity by writing speed differential, DC1 for complex similarity by writing velocity and speed, DC2 for complex similarity by writing velocity and speed differential, DC3 for complex similarity by writing velocity, speed and speed differential.
(5)
The signature verification results with complex similarity are shown in Fig. 6-8. Fig. 6 shows signature verification result with complex similarity by writing velocity and speed. The best EER result records 8.33% when α = 0.7 and β = 0.3. Fig. 7 shows signature verification result with complex similarity by writing velocity and speed differential. The best EER result records 8.22% when α = 0.45 and β = 0.55. Fig. 8 shows signature verification result with complex similarity by writing velocity, speed and speed differential. The best EER result records 8.18% when α= 0.4, β = 0.1, γ = 0.5 or α = 0.5, β = 0.2, γ = 0.3. Furthermore, experiments when α = 0.6, β = 0.2, γ = 0.2 and α = 0.7, β = 0.2, γ = 0.1 records closely best EER result.
The best signature verification result was calculated as mentioned before. To use the system practical, we must use the threshold which has been fixed. Therefore, we calculate the verification rate which is using common threshold for each eight experiment. In this case, we set the fixed threshold as average of suitable threshold for each experiment. Fig. 9 shows result of signature verification. Similarity of signature will be calculated as following expression.
(6)
As a result, the signature verification method which records best performance was method that uses complex similarity by writing velocity and speed differential. The signature verification method records 10.29% as average FRR, 9.72% as average FAR and 9.96% as total error rate.
6. Conclusions
In this paper, we proposed to change self verification method as signature verification from verification method which use password at existing multiuser online kanji learning system. By adopting signature verification, we succeed to improve reliability and usability of the system. From experimental result, we adopted signature verification to use writing velocity and speed differential as feature. We adopted it because there was no big difference between the verification rate which records lowest EER such as 8.18% and it does not need to calculate the similarity by writing speed so the verification time will be decrease. Average EER of the adopting method was 8.22% when we establish threshold most suitable. On the other hand, average of FRR records 10.29%, FAR records 9.72% and total error rate records 9.96% when we establish threshold as constant, to simulate if the signature verification function is implement in actual running system. From mentioned above, we could say new multiuser online kanji learning system become more practical when we compare with existing system.
As future works, we could consider following improvements to realize the widely use of the system. First, reinforcement of the communication tools between server and client to realize easy and deep understanding at kanji learning. Second, addition of the function that could statistically compute the tendency of each user’s weak points at writing kanji, radical, and so on. The addition of this function will realize more effective kanji
learning. Finally we perform the improvement of signature verification method. This will realize construction of more reliable multiuser online system.
References
[1] R. Tatsuoka, M. Yoshimura, Development of a kanji learning system for foreign students or elementary students, IEICE Trans. & System, ET96-36, June 1996.
[2] N. Takesue, K. Mochida, A. Kitadai, M. Nakagawa, A handwriting-based kanji learning system enabling teachers to designate evaluation points, IPSJ SIG Technical Reports (15) (2005) 15-22.
[3] T. Iwasaki, A. Kitadai, J. Tokuno, M. Nakagawa, A web based training system to learn kanji, IPSJ SIG Technical Reports 16 (2006) 17-24.
[4] A. Takeda, J. Shin, Character Learning System Using Inter-stroke Information, in: Proceedings of the 8th International Conference on Knowledge-Based Intelligent Information, Wellington, New Zealand, Sep. 2004, pp. 165-174.
[5] Y. Shimizu, J. Shin, User friendly kanji learning system and its evaluation, Master’s Thesis, University of Aizu, Mar. 2010.
[6] J. Takahashi, Status of primary schoolchild’s keyboard skill through development of keyboard input study web site (in Japanese), Univ. of Toyama Information Technology Center Public and Information, 2004, Vol. 1, pp. 16-19.
[7] S. Nakajima, T. Hamamoto, S. Hangai, On-line signature verification using pen inclination, IEICE (1998) 15-22.
[8] B. Yanikoglu, A. Kholmatov, Online signature verification using fourier descriptors, EURASIP Journal on Advances in Signal Processing 2009 (2009).
[9] E. Maiorana, P. Campisi, A. Neri, Template protection for dynamic time warping based biometric signature authentication, in: 16th International Conference on Digital Signal Processing (DSP 2009), Santorini, Greece.
[10] W.-D. Chang, J. Shin, DPW approach for random forgery problem in online handwritten signature verification, in: 4th International Conference on Networked Computing and Advanced Information Management, Korea Republic, 2008.
[11] D.-Y. Yeung, H. Chang, Y. Xiong, S. George, R. Kashi, T. Matsumoto, G. Rigoll, SVC2004: First international signature verification competition, ICBA, July 2004.