Keystroke recognition is a behavioral biometric which authenticates (verifies the claimed identity) or identifies an individual (recognizes a valid user) based on their unique typing rhythm. Unlike physiological biometrics, such as fingerprint or iris, where specialized sensors are necessary to collect data, keystroke biometrics utilizes off-theshelf physical keyboards or virtual keyboards in smartphones or PDAs. Thus, keystroke recognition offers a low-cost authentication and is easily deployed in a variety of scenarios. Two events constitute a keystroke event: key down and key up. The key down occurs when the typist presses a key. The key up is associated with the event that occurs when the pressed key is released. Using these two events, a set of intra-key and inter-key features commonly called hold times, delay times and key down-key down times can be extracted. Hold times constitute the finite amount of time a particular key is pressed. Delay times constitute the latency between the release of the current key and the pressing of the next key. Delay times may be negative given that individuals can press the next key prior to releasing the current key. Finally, key down-key down time represents the finite amount of time between successive key down events. While other keystroke events can be extracted (such as key pressure and location of a particular Shift key and special keys such as Alt, Ctrl, etc.), these features are not commonly used as their collection from commercially available keyboards is not straightforward. Keystroke data can come either from a predetermined body of text (fixed-text analysis) or from any unrestricted text (free-text analysis). Fixed-text analysis from strings similar to passwords can be used for one-time authentication. On the other hand, free-text analysis can allow more complex usage scenarios. Given sufficient corpus of user activity, it can support continual user authentication and ensure that an impostor has not gained control of the device. Similar to other behavioral biometrics, keystroke recognition exhibits a phase in which the individual learns the motor skills required to enter text entries. For the fixed text, habituation phase includes the cognitive activity related to string memorization as well as the adaptation of motor

In this paper we explore the possibility of examining an iris image and identifying the sensor that was used to acquire it. This is accomplished based on a classical pixel non-uniformity (PNU) noise analysis of the iris sensor. For each iris sensor, a noise reference pattern is generated and subsequently correlated with noise residuals extracted from iris images. We conduct experiments using data from seven iris databases, viz., West Virginia University (WVU) non-ideal, WVU off-angle, Iris Challenge Evaluation (ICE) 1.0, CASIAv2-Device1, CASIAv2-Device2, CASIAv3 interval, and CASIAv3 lamp. Results indicate that iris sensor identification using PNU noise is very encouraging, with rank-1 identification rates ranging from 86%-99% for unit level testing (distinguishing sensors from the same vendor) and 81%-96% for the combination of brand (distinguishing sensors from different vendors) and unit level testing. Our analysis also suggests that in many cases, sensor identification can be performed even with a limited number of training images. We also observe that JPEG compression degrades identification performance, specifically at the sensor unit level.

As large scale biometric systems have been deployed throughout the landscape of the federal government, a new set of challenges and expectations have arisen. The ability to address increasing database sizes, demands on throughput, and privacy / legal issues are just a few examples of challenges at the forefront. Despite these challenges, the large scale systems are expected to continue to improve in terms of accuracy. In response to such challenges and expectations, systems relying on a single source of biometric input are transitioning to multi-biometric implementations. Traditional motivations behind the application of multi-biometric systems have centered on improving recognition performance, increasing population coverage, deterring spoof attacks, and reducing failure-to-enroll rates. The potential for improvement associated with these motivations is typically thought to come at the expense of increased processing time and computational complexity. In this paper, we provide supporting arguments for expanding the basis of multi-biometric system evaluation to reflect a holistic scope. Namely, the best multi-biometrics systems will maximize the degree in which requirements corresponding to the characteristics of universality, uniqueness, permanence, measurability, performance, acceptability, and circumvention are met. Additionally, we will demonstrate how evaluating the requirements of large scale federal multi-biometric systems necessitate an increased level of granularity beyond the traditional seven characteristics. Finally, we provide an example prioritization scheme which demonstrates how the overall measure of success need not uniformly consider all characteristics.

Traditionally, a chain of evidence or chain of custody refers to the chronological documentation, or paper trail, showing the seizure, custody, control, transfer, analysis, and disposition of evidence, physical or electronic. Whether in the criminal justice system, military applications, or natural disasters, ensuring the accuracy and integrity of such chains is of paramount importance. Intentional or unintentional alteration, tampering, or fabrication of digital evidence can lead to undesirable effects. We find despite the consequences at stake, historically, no unique protocol or standardized procedure exists for establishing such chains. Current practices rely on traditional paper trails and handwritten signatures as the foundation of chains of evidence. Copying, fabricating or deleting electronic data is easier than ever and establishing equivalent digital chains of evidence has become both necessary and desirable. We propose to consider a chain of digital evidence as a multi-component validation problem. It ensures the security of access control, confidentiality, integrity, and non-repudiation of origin. Our framework, includes techniques from cryptography, keystroke analysis, digital watermarking, and hardware source identification. The work offers contributions to many of the fields used in the formation of the framework. Related to biometric watermarking, we provide a means for watermarking iris images without significantly impacting biometric performance. Specific to hardware fingerprinting, we establish the ability to verify the source of an image captured by biometric sensing devices such as fingerprint sensors and iris cameras. Related to keystroke dynamics, we establish that user stimulus familiarity is a driver of classification performance. Finally, example applications of the framework are demonstrated with data collected in crime scene investigations, people screening activities at port of entries, naval maritime interdiction operations, and mass fatality incident disaster responses.

Arguably the most important task in iris recognition systems involves localization of the iris region of interest, a process known as iris segmentation. Research has found that segmentation results are a dominant factor that drives iris recognition matching performance. This work proposes techniques based on probabilistic intensity features and geometric features to arrive at scores indicating the success of both pupil and iris segmentation. The technique is fully automated and therefore requires no human supervision or manual evaluation. This work also presents a machine learning approach which utilizes the pupil and iris scores to arrive at an overall iris segmentation result prediction. We test the techniques using two iris segmentation algorithms of varying performance on two publicly available iris datasets. Our analysis shows that the approach is capable of arriving at segmentation scores suitable for predicting both the success and failure of pupil or iris segmentation. The proposed machine learning approach achieves an average classification accuracy of 98.45% across the four combinations of algorithms and datasets tested when predicting overall segmentation results. Finally, we present one potential application of the technique specific to iris match score performance and outline many other potential uses for the algorithm.

In this paper we study the application of hardware fingerprinting based on PRNU noise analysis of biometric fingerprint devices for sensor identification. For each fingerprint sensor, a noise reference pattern is generated and subsequently correlated with noise residuals extracted from test images. We experiment on three different databases including a total of 20 fingerprint sensors. Our results indicate that fingerprint sensor identification at unit level is attainable with promising prospects. Our analysis indicates that in many cases identification can be performed even when one only has access to a limited number of samples. For two of the three databases one can train on less than 8 images per device and establish sensor identification with little or no misclassification error. On the third database, high levels of identification performance can be achieved when training on similar amounts of images required for other types of sensor identification such as cameras or scanners.

Keystroke dynamics are becoming a well-known method for strengthening username- and password-based credential sets. The familiarity and ease of use of these traditional authentication schemes combined with the increased trustworthiness associated with biometrics makes them prime candidates for application in many web-based scenarios. Our keystroke dynamics system uses Breiman's random forests algorithm to classify keystroke input sequences as genuine or imposter. The system is capable of operating at various points on a traditional ROC curve depending on application-specific security needs. As a username/password authentication scheme, our approach decreases the system penetration rate associated with compromised passwords up to 99.15%. Beyond presenting results demonstrating the credential hardening effect of our scheme, we look into the notion that a user's familiarity to components of a credential set can non-trivially impact error rates.

In this paper we describe decision dependability theory for binary classification problems. Each classification decision is subjected to an informal evaluation of confidence that can be attributed to it. The confidence measure emanates from the strength of evidence which supports the decision. We utilize decision dependability theory in the context of multimodal biometric identity verification systems. The confidence measure is estimated from quality and match scores of biometric samples using subjective Bayesian methodology. We demonstrate how fusion algorithms, such as Bayesian belief networks and likelihood ratio, can be complemented with decision dependability in order to predict classification errors. Furthermore, we illustrate how decision dependability can be used to rectify incorrect decisions.