INTRODUCTION:

The main purpose of XSEC project is to develop a user authentication and identification system that works upon biological features of the user and delivers high precision to be used for different security purposes. User authentication is a binary decision system that decides whether a supplied feature belongs to the claimed person or an impostor tries to log in. Obviously, acceptance of an impostor to the system is completely intolerable as a security system is the case. On the other hand frequent rejection of genuine users may be disturbing. So the system should be able to clearly differentiate biological characteristics among different people while compensating in-person differences. System is aimed to have complete robustness against channel transformations and acceptable noise levels so to be able to work in practical applications with an error rate close to zero. 
     Biometric methods of identification are currently being used to replace the less secure ID/Password method of user authentication, that is, verifying that people are who they say they are. Using biometric identifiers for personal authentication reduces or eliminates reliance on tokens we must carry with us, or the arcane strings of letters and numbers we are forced to memorize. Tokens, such as smart cards, magnetic stripe cards, and physical keys can be lost, stolen, or duplicated. Human memory is notoriously unreliable; according to recent estimates, at least 40% of all help desk calls are password or PIN-related. Losses attributed to fraud, identity theft, and cyber vandalism due to password reliance run well into the billions. Although passwords have traditionally been used for personal authentication, they have nothing to do with a person's actual identity!
     The best method for a particular solution depends on the type of users, desired accuracy, desired transaction speed, cost parameters and any cultural issues or sensitivities. Technology today provides us tools like Fingerprint characterization, Hand geometry (palm print) characterization, Voice pattern characterization, Iris pattern characterization, Retinal patterns characterization, Facial feature characterization etc.
      XSEC has chosen facial feature characterization and voice feature characterization to identify people.

XSEC ARCHITECTURE:

XSEC authenticates persons based upon there voice and face characteristics. Whole architecture is divided into two different modules that work independently upon speech and face characterization respectively. Block diagram shown below gives a rough idea about the functionality of the System.

 
 

Face characterization module takes input from a digital camera at Point-1, while voice characterization module takes input speech sample from a microphone at point-2.

SYSTEMS OVERVIEW:

Voice Characterization module

Speaker recognition systems are grouped into two categories that are generally called speaker verification and identification respectively. Speaker verification is the determination from a voice sample if a person is who he or she claims to be. On the other hand speaker identification determines which one of a group of known voices best matches the input voice sample.
     Another classification for voice recognition systems takes the text dependence properties of the system into account. Text dependent systems are those for which the speech used to train and test the system is constrained to the same word or phrase. On the other hand for text independent systems training and testing speech are completely unconstrained.
     Roughly speaking, a factor that constitutes the information content of spoken utterances is speaker characteristics + spoken phrase + emotions and additional noise, channel transformations etc. For speaker recognition case, our concern is how to distill speaker dependent characteristics from all others. So we may refer to speaker verification as a three-state problem. 

1. First step is to extract speaker dependent features from spoken utterances. For this purpose different feature sets have been proposed to reflect speaker’s identity best most famous of which are listed below...

·Mel-frequency Cepstrum coefficients

·Liner prediction Cepstrum coefficients

·Log-Cepstrum coefficients 

2. Second stage is to build a statistical model that will successfully reflect the characterization of the chosen feature set. Below is the list of popular statistical models that are widely used and proved to be successful.

·Vector Quantization

·Gaussian Mixture Models

·Hidden Markov Models

·Neural Network Architectures 

3.  The third step which is the decision making step is where we compare he input voice to the claimed speaker model and make a decision about the identity of the speaker.
    XSEC uses Mel-frequency Cepstrum coefficients method to extract peoples’ voice characteristics with techniques like RASTA filtering and cepstral mean normalization to compensate the performance degradation due to handset mismatches.
   XSEC implements a unique technique for handset identification and normalization for the handset differences. For statistical modeling XSEC uses hybrid architecture consists of HMM and ANN. Hence, it gives higher accuracy than its other competitors in the market.

 Face Characterization module

Along with voice identification XSEC realizes a face authentication system. Face recognition functionality can be divided into two sub-parts, modeling of person’s image followed by the recognition phase. Broadly speaking, face recognition problem can be solved by using HMM (Hidden Markov Modeling) or with Eigen Faces. XPEG uses its own technique that makes several compromises between two methods stated above and delivers higher recognition efficiency.

BACKBONE OF XSEC ENGINE:

Faceprint & Voiceprint database: User is required to give his voice sample and few training face samples when he accesses the system for the first time, i.e. Training session. Supplied faces are in a prescribed face orientation and spoken utterance is a predefined text of acceptable length that successfully reflects the vocal properties of the language and the user. Once the speech samples are taken from the user, it is processed and a property matrix that reflects the user’s voice is saved in the database. This property matrix is technically called the voiceprint. Similarly, system computes to get a property matrix corresponding to the persons’ face and is stored in the database, which is technically called faceprint.  So the data kept in the database is not the speech sample or picture but the voiceprint and faceprint of the user. 

Login phrase: It is vital to practicality of the system that required login phrase duration should be as short as possible. XSEC works with login phrases of 1 picture and 3-4 words randomly generated and flashed to the user.

Decision-making: XSEC tries to determine user’s identity by face recognition methods. Once user identity is determined XSEC uses voice authentication methods to verify this identity. User needs to supply his spoken utterance that will be used for authentication and the claimed speaker information is available to the system, the feature matrix is extracted from the speech data and compared to the voiceprint of the claimed speaker. If the similarity score exceeds a predefined threshold value, the user is accepted.

BIOMETRIC FEATURE SECURITY:

If a criminal steals or guesses the password, it is very easy to have it changed. There is a fear, however, that if a criminal gets hold of a biometric template, the damage is irreparable - there is no way to change that part of your body. XSEC-2.1 works in a network environment and keeps all the user specific data on a secured server computer. Only way to introduce a new user template is through XSEC-2.1 client, hence template manipulation chances are well taken care of.

DRIVERS FOR XSEC PROJECT

There are four main benefits to use biological verification:

1.Improve Security: This is the key objective of using voice and face biometrics to improve the security of sensitive information and reducing fraud. A password is what people know. But the voiceprint and face-print is what they have.
2.Reduce Costs: Using an automated authentication process reduces salary expenses, toll-costs and call hold times and allows live agents to focus on revenue-generating calls.
3.Improve Service: Speaker verification combined with speech recognition (rather than touchtone) allows callers to interact naturally with the application, so callers can actually enjoy the automated experience. The combination of speech recognition and speaker verification makes it possible to simultaneously identify AND authenticate callers in one step, allowing them to simply speak an Account ID or their name to access the account and be authenticated.
4.Save Time: Whereas standard authentication questions and associated hold times may take an average of 80 seconds, speaker verification happens instantaneously, especially when the voiceprint and faceprint is associated with the Account ID, removing the time needed for a separate password.

MARKET OPPORTUNITY

As enterprises and organizations increasingly turn to biometrics to ensure their customers’ safety, the market opportunities for voice/face authentication are infinite.

Market	Application	Drivers
Financial Services	Access to Banking, Brokerage,	Reduce Financial Risk
Telecom	Call Center Applications Unified Messaging Auto Attendant	Reduce Fraud Protect Personal Information Competitive Advantage
Retail	Order Entry Personalized Service	Reduce Fraud Increase Revenue
Enterprise and IT	Access to Intranet, Extranet and Corporate Applications	Increase Security Reduce Cost
Travel	Frequent Customer Services	Convenience Personalization
Internet	Authenticate Users for Internet Banking and e-Commerce	Reduce Financial Risk
Hospitals,Insurance	Access to Patient Information Authorize Drug Prescription Authorize Insurance Payment	Protect Personal Privacy Reduce Fraud
Government/Military	Access to Sensitive Information Parolee Tracking	Increase Security Reduce Cost