Continuous Biometric Verification

If a protected resource needs to be continuously monitored to ensure that an impostor does not attempt to access the resource at any point in time, one time verification is not sufficient.  In this work, we utilize the face biometric and video data to monitor a subject passively.  Without cooperation from the subject, passively monitoring requires significant amount of tolerance for intra-class variations.  In order to account for the large intra-class variations without requiring a large amount training data we utilize a 3D model of the subject to track and recognize the subject at large pose variations.  Data from previous frames is incorportated in a probabilistic framework to provide a confidence measure at any point in time that the genuine user is still using the protected resource.  If the confidence falls below a threshold, the user is asked to provide a frontal face image to perform accurate verfication.

Human-like Similarity Measure for Face Recognition

In order to capture how humans are able to recognize faces, we developed a system that mimiced the human similarity measure.  A neural network was trained on data that consisted of the border (manifold) of the face space where a human expert would classify a subject as genuine or impostor.  This data was generated by warping a genuine image of a subject toward many other subjects.  Genuine images were created by warping until a threshold where the result is still recognizable as the genuine subject.  Impostor images were created by warping until the result was a 'look-alike' of the person, but no longer recognizable as the genuine subject.

Local and Global Fusion for Signature Verification

Many types of features and their corresponding matchers exist for signature verification.  Many of the algorithms can be seperated into two types; global and local methods.  In this work, we investigated the performance of combining a local matcher (string matching) and a global matcher (hidden markov model).  The results were able to show that each matcher is able to extract some independent information and the combination of the two produced an improved result.

Signature and Voice Authentication for Tablet PC

Many medical professionals are starting to utlize the advantages of a table PC to access medical records.  While this provides many advantages to the user, it also emits concern for the privacy of the medical records.  HIPPA regulations and public concern create the need for a reliable and secure form of authentication.  We propose a system that utilizes the fusion of signature and voice biometrics to authenticate the user of a tablet PC.  This type of system is easily incorporated into a tablet PC because most come built in with a stylus and an internal microphone.  These forms of authentication are also well accepted by the public.  A prototype system was created and tested on a database of 100 users.  Results proved that this type of authentication is sufficient to protect medical records while allowing medical professionals easy access.