| Project | Short Descrption | Impact |
|---|---|---|
| Positive Personal Identification (PPID) | The key requirements in a largescale benefit or voting enrollment system requires human enrollment that can’t be fooled to takin duplicate entries. Many government systems have this basic need. In our PPID system, fingerprint based duplicate checking was deployed using a scalable system. The key innovation was around a sub-linear search algorithm for fingerprint matching using minutia-triple in the lines of geometric hashing. | Code delivery, Papers and IP |
| Large scale fingerprint indexing | A follow-on FOAK years later was approved to improve the system designed through PPID. We introduced many textural features around minutia to improve the system performance. As a result, a novel registration free method of fingerprint verification was also developed. | Code delivery, Papers and IP |
| Fusion of Biometrics and Biography | Often in large biometrics enrollment systems (e.g., DMV, passport, visa) a large amount of biographical details is available. In fact, the biographical matching is also quite in exact as names are spelt differently and addresses are written differently. In addition to fusion of several biometrics modalities, we introduced inclusion of biographic engine in the mix of biometrics to extend the utility of biographic information into the de-duplication process. | Code delivery, Papers and IP |
| Bodycam based analytics | It is perceived that the ultimate truth can be collected in videos in many law enforcement encounters. The bodycam and similar devices used in civil applications as well collect video that is often very noisy due to the environment they operate. We developed human detection and processing algorithms including face and human attributes from video frames for indexing and recognition. A research issues in this area is how to protect the privacy of entities such as face, cars and other identifiable information. Using blockchain technology, we have show how multiple version of the video can be generated while ensuring that no tampering happened to the core video. | Code delivery, Papers and IP |
| Face recognition in video and Face analytics as services API | Face is ubiquitous in almost all videos. Many applications need matching faces and classify different facial attributes beyond typical law enforcement use. In fact, most deep learning methods get applied invariably to face analytics. We have shown how to learn a face embedding network based on small learning dataset based on the physics of the problem. Once a reliable embedding is available, we can use it for several attribute prediction including gender, age, and expression to name a few. These innovations made into API based services for internal and external clients. | Code delivery, Papers and IP |
| Smarter Planet: analytic integration and better the news | In a complex system involving multiple analytics it is not known how to make the outcome consumable to the use. Based on a basic sports videos, we analyzed different models to understand how to interpret starting line-up differences and player identification. In social media, often the crowd sourcing of information makes us possible to understand the pattern of a story based on how quickly it becomes viral by spreading in the social networks and thus almost creating and substantiating news items that can be created by new agencies. The multi-modal perception helps validate the content as well as shows us the trend. This is an external government funded project. | Code delivery and IP |
| Confidence interval estimation | In a pattern recognition system, performance reporting is fundamental way to describe how the system behaves. Often ROC has been used for this purpose. However, without a confidence interval around the error measurements, it is not meaningful. We reported a non-parametric bootstrap based technique called subset bootstrap to compute confidence interval in biometrics system performance reporting. | Papers and standard |
| eSentry | Secure authentication: Biometrics attack model and spoofing: Authentication of identity must be secure in all applications. Biometrics systems need to defend against many attacks starting with presentation attacks. We developed a threat model for biometrics systems highlighting the dangers that the designers need to handle. In the process, the template secure storage is an important component. In a personal smartcard, one can store the template and often the parts of matching can be carried out in a collaborative fashion to provide security. An extension of the similar approach can be to match in a secure environment using hardware secure modules. However, if the presentation attacks are not detected using novel methods, the backend security can be rendered useless. We have proposed techniques for replay attack detection by using data hiding methods and face and iris presentation attacks using 3-d sensing from 2-d video. | Papers, Standards and IP |
| Cancelable biometrics | A biometrics measurement while provides a non-repudiable connection to the person behind a transaction, often it can be misused violating privacy of the person. For example, two databases can be matched and links established using biometrics since it is a unique descriptor of the person. A second issue related to the non-repudiablity is that if the data is lost due to mishandling the data can be replaced. In order to address these two issues, we introduced cancelable biometrics by transforming the original biometrics using a non-invertible, repeatable transform during enrolment. During verification the same transform is applied to the input to match against the transformed template. Since each database can use a different transform, matching across databases becomes difficult and the enrollment data can be revoked by reenrolling the person with a new transform. This technique is popular even today in the biometrics research community first introduced by us decades ago. | Papers, IP and Media coverage |
| Trusted secure authentication on mobile devices (Usable Multi-Factor Authentication and Risk-Based Authorization) | ABiometrics authentication on mobile devices is an interesting paradigm since the mobile devices house a set of sophisticated sensors that are typically not associated with a biometrics sensor. For example, GPS, proximity sensor, altitude sensor, accelerometer can provide context that can help with establishing the transaction environment. Thus, suitable biometrics can be selected for authentication depending on the situations. Along with different biometrics modalities, contextual sensing can be added to define the authentication policy. These techniques initiated from our FOAK project entitled “TSAM” ended up being useful in a DHS funded project for establishing financial industry secure transaction methods. | Code delivery, Papers and IP |
| Anonymous Biometrics Matching | Many a times, biometrics search methods needs not be linked the persons behind the biometrics until needed. For example, the airline industry needs match the passengers against a law-enforcement watch list. However, the law-enforcement agencies can not reveal the watchlist to airlines. In order to facilitate a solution for this, we proposed anonymous biometrics matching based on our cancelable biometrics. With the watchlist available as a transformed version, the airline passenger biometrics can be transformed by a service provider and match against the watchlist. When there is a need to go behind strong matches from the list, the law enforcement can go back to the airlines to reveal the persons of interest. Thus the overall privacy of the travelers is not compromised. | papers and IP |
| Fully Homomorphic encryption | AWe believe that if computations can be carried out on encrypted domain, there will less concerns about sharing data across many domains where data is protected by local compliance regulations. In order to facilitate learning in encrypted domain, we have shown that FHE can be deployed to produce near similar results. However, this is a baby step towards addressing the larger machine learning issues in encrypted domain. We continue to evolve our research in this area. | Papers and IP |