Skip to main content

Fast Correlation Filtering for Biometric Recognition

Researcher: Vijayakumar Bhagavatula

Research Area: Security of Cyber-Physical Systems

Cross Cutting Thrusts: Cryptography


Correlation filter-based biometric recognition has performed very well in biometric recognition challenges such as face recognition grand challenge (FRGC) and iris challenge evaluation (ICE). In this approach, during the enrollment stage, multiple biometric images (in this example, face images) of a subject are collected. The 2D Fourier transforms (FTs) of these images are computed using 2D fast Fourier transform (FFT). These FFTs are used to design a complex-valued correlation filter which is used during the recognition stage. During the recognition stage, the 2D FFT of the live biometric image is computed and it is multiplied by the correlation filter. The 2D inverse FFT of this product yields a 2D correlation output. Presence of a sharp peak in the correlation output indicates a match (i.e., authentic) and the absence of such a peak means the input does not match (i.e., an impostor). The correlation peak sharpness is quantified by metrics such as peak to correlation energy (PCE). For 1:N matching, the template yielding the largest PCE value is chosen as representing the subject’s identity. In this project, we investigate several approaches to speed up correlation filter-based biometric recognition. Methods include: using 1D projection, avoid FFTs, quantifying filter phase and using transforms other than Fourier transforms.

Improved Biometric Encryption using Correlation Filters

The username/password paradigm is the most commonly used authentication mechanism for accessing physical devices and spaces. However, passwords have the well-known limitations that they are often not as strong as they need to be and can be compromised via dictionary attacks. One way to improve the security is by using biometric signatures (e.g., face image, iris image, etc.) to verify the identity of a claimant. The main challenge in biometric recognition is the inherent variability in biometric signatures, e.g., the appearance of a face image can change significantly with illuminations, pose, expressions, etc. We have developed correlation filter methods to reduce the error rates in the presence of such variability. A more common mechanism for authenticating a user is through cryptographic keys. While such cryptographic keys exhibit sufficient entropy (because of their inherent randomness), they are usually stored in and retrieved from a physical medium (e.g., smart card), most often using a password. Thus, they are only as secure as the password being used. One way to combine the strengths of biometrics and cryptography to generate encryption keys from the biometric signatures. Of course, if not designed carefully, the combination may inherit only the weaknesses of the two. The goal of this project is to improve the performance of biometric encryption schemes can be improved by using advanced correlation filters in place of conventional methods, e.g., those based on minutiae (for fingerprints) or iris codes (for iris images). Correlation filters offer advantages of shift-invariance, graceful degradation and distortion tolerance, which should prove very useful in handling the normal variability anticipated in biometric signatures. We have demonstrated the superior matching performance of advanced correlation filters in face recognition grand challenge (FRGC) and iris challenge evaluation (ICE).