Naohisa Komatsu

Digital watermarking based on process of speech production

A speech production procedure can be divided into three parts, namely the glottal source, articulation and radiation, respectively. We propose a watermarking method for speech by manipulating the articulation in the process of speech production. We apply our method to CS-ACELP(G.729 standard), which is the ITU-T approved recommendation. It provides a low bit rate 8 kb/s speech coding algorithm with wire/line quality. The watermarked vocal tract model is expressed by codebooks made by LSP(Line Spectrum Pair) parameters. The codebook vectors replace some of the extracted LSP. Speech is synthesized using replaced LSP. We generate a couple of codebooks using a unique method to modify the LSP of the spectrum envelope. Shortening the width of the LSPs creates one watermarked codebook, and the second codebook is created by stretching the LSP of both sides of each formant. There are ten LSP dimensions in each voice frame of the CS-ACELP decoder. In the detecting process, the weighted Euclidean distance(WED) between the watermarked codebooks and the extracted LSP will be calculated. Whether the watermark is embedded will be judged by utilizing the calculated WED. Evaluation tests on detection accuracy will be discussed with simulation results.

A gait recognition method using HMM

New biometrics identification methods, including person identification using a walk characteristic (gait recognition) is attracting attention. In this paper, an identification method on gait recognition is proposed. In the proposed method, outlines of person are represented by P-style Fourier descriptions, and these time-serial walking characteristics use HMM. It is shown that the proposed method brings the high rate of identification and is effective in indoor environment with simulation results.

Speaker recognition using kernel mutual subspace method

We propose a method of person authentication based on lip movement by using the kernel mutual subspace method. Algorithms for person authentication using lip movements have already been developed, but in case that the distribution of lip images has a nonlinear structure, the identification accuracy is degraded because the conventional recognition methods do not assume such nonlinearity. We have investigated the distributions of lip images in feature space and experimentally obtained their nonlinear properties. To solve the problem of nonlinearity, we propose a labiate person authentication method based on the kernel mutual subspace method. We evaluated accuracy and stability of the proposed method with demonstration experiments results.

Text-indicated writer verification using hidden Markov models

We propose an HMM-based text-indicated writer verification method, which is based on a challenge and response type of authentication process. In this method, a different text including ordinary characters is used on every occasion of verification. This text can be selected automatically by the verification system so as to reflect a specific writers personal features. The specific writer is accepted only when the same text as indicated by the verification system is inputted, and the system can verify the writers personal features from the inputted text. Moreover, the characters used in the verification process can be different from those in the enrollment process. This method makes it more difficult to get away with forged handwriting than the previous methods using only signatures. In the proposed method, the characteristics of the indicated text and each writers personal features are both represented by using hidden Markov models.

A Motion Detection Method on Video Image by Using Hierarchical Pixels

It is not easy to extract significant information from image data concerning the motion and the structure of an object, but this technique is indispensable to the future development of video coding and recognition of a moving object. The block matching method [1] is employed widely as a motion detection method in the motion-compensated interframe coding. However, it has a problem in that the motion is estimated by the prediction error power minimization condition, which may lead to an instability in accurate motion detection. This paper proposes a new block matching method that has the following features: the motion is estimated and detected by stepwise switching from larger block size to smaller block size; by this scheme, the motion detection accuracy and the spatial resolution can be improved, and the processing complexity can also be reduced without degrading the detection accuracy.

A study on vulnerability in on-line writer verification system

The analysis of vulnerabilities and threats in biometrics based personal authentication systems is indispensable for the development and promotion of wide spread utilization of biometric technologies in accordance with international standardization of information security technologies. In this paper, we discuss the vulnerabilities and threats in a writer verification system which is an application of biometric technologies using handwriting information. We propose some attack methods that use verification results or user templates of a writer verification system. From some simulation results, it is clear that an attacker can estimate an authorized users biometric information (handwriting information) and produce some forged handwritings which are accepted by the verification system.

A proposal for a text-indicated writer verification method

We propose an on-line writer verification method to improve the reliability of verifying a specific system user. In the proposed method, a different text including ordinary characters is used on every verification process. This text can be selected automatically by the verification system so as to reflect the specific writers features. A specific writer is accepted only when the same text, which is indicated by the verification system, is written, and the system can verify the writers personal features from the written text. The proposed method makes it more difficult to disguise writer himself with forged handwriting data than the previous methods using only signatures.

Extraction of personal features from stroke shape, writing pressure and pen inclination in ordinary characters

We propose an extraction method of personal features based on online handwritten characters including writing pressure and pen inclination information. In the proposed method, any handwritten character (i.e., ordinary character) is described by a set of three dimensional curves, and personal features are described by a set of Fourier descriptors for the three dimensional curves. From some simulation results using handwritten data, it is clear that the proposed method effectively extracts personal features from ordinary characters.

Internet Traffic Classification Using Score Level Fusion of Multiple Classifier

Internet traffic is continuously growing fast due to the rapid spread of the internet and the speed-up of the internet connection. Also, the applications provided on the internet have become more diversified. To support the QoS requirements for these internet applications, it would be better to measure the traffic volume according to the applications. Therefore, we are engaged in the application classification method, which is an offline technique for identifying the applications in units of flow. In some application classification methods, the applications of the target flows are analyzed according to their statistics on traffic metric, or features. We focus on these feature based classification methods, since the methods have the advantage that the port number and the packet payload need not be checked for classification. In the field of the machine learning, the classification methods that consist of multiple classifiers have been discussed. This is why the classification methods are improved in performance. However, the conventional feature based classification methods consists of single classifier. Also, the design of multiple classifiers has hardly been discussed. The design includes the way of combining some classifiers. Here, we introduce the fusion of multiple classifiers and propose applying the score level fusion using feature vectors to concatenate each classifier score to classify the flow into applications.

Multimodal Biometrics of Lip Movements and Voice using Kernel Fisher Discriminant Analysis

We propose a method of text-prompted speaker recognition based on multimodal biometrics by using the kernel Fisher discriminant analysis. Algorithms for person authentication using multimodal biometrics have already been developed. When the decision boundary between genuine and imposter has a nonlinear structure, the identification accuracy will be degraded because the conventional recognition methods do not assume nonlinearity of decision boundary. We have investigated the score distribution. We also need to control the security level (FAR,FRR) by the threshold. To solve the problems, we have introduced a method of speaker recognition based on multimodal biometrics by using the kernel Fisher discriminant analysis. We experimentally demonstrate the proposed methods effectiveness with simulation results and show that the method achieved high accuracy