Novel methods for fingerprint image analysis detect fake fingers

Biometric systems are an exciting emerging technology. They allow us to forget our passwords or where we left our keys because the information required to login or unlock something is encoded in our anatomy and behavior. However, security problems arise when we display our passwords or keys, or when we deposit our fingerprints on the side of a drinking glass. Each of these scenarios leaves us open to attack. Spoofing is an attack at the sensor level in which a biometric sample is replaced by an imposter's sample. Recent research demonstrates that it is possible to spoof or deceive a variety of fingerprint scanners using simple molds made from plastic, clay, Play-Doh, silicone, or gelatin materials.^1,2 Figure 1 compares live and spoof fingerprint images.

To combat spoofing, methods of liveness detection measure physiological signs to ensure that only live fingers are captured for enrollment or authentication. Previous efforts used medical-based hardware measurements, such as from pulse oximetry, temperature, and odor. Our software-based approach analyzes the patterns in a fingerprint image relating to measurable characteristics that differ between spoof materials and live fingers.

Figure 1. Images of typical (a) live and (b) spoof fingerprints.

Although other software-based liveness approaches evaluate skin deformation,^3,4 our group focuses on the perspiration pattern along fingerprint ridges and the noise pattern along valleys.^5–7 Figure 2 illustrates ridge and valley extraction from a live image. Live fingers, unlike spoof and cadaver fingers, have a distinctive spatial perspiration pattern when in physical contact with the capturing surface of a fingerprint scanner. Spoof fingers have a distinctive noise distribution—unlike the clean valley structure of live fingers—due to the spoofing material's properties when placed on a scanner. Both of these approaches to liveness detection use the same algorithmic structure. The first step checks image quality and cleans signal strength. Then the region of interest is extracted, which involves the core of the fingerprint, and ridge and valley segmentation. Ridge and valley signals, wavelet coefficients, and an intensity histogram contribute to the creation of a feature set. The final step classifies the image.

Figure 2. Extracted (a) ridge and (b) valley structures from a live fingerprint image.

Live fingers demonstrate both dynamic and static moisture patterns along the ridges due to perspiration and pores. However, spoof fingers have greater noise along the valleys because the materials are transmuted easily by human pressure and because granules collect along the valleys due to the properties of the materials. Image and signal processing, and pattern-recognition algorithms can quantify both moisture and noise using wavelet and statistical approaches. The ridge perspiration detection method achieves an equal error rate (ERR) of approximately 10% when distinguishing live from spoof images in a data set of 58 live, 80 spoof, and 25 cadaver images.^5,6 The valley noise detection method⁷ achieves an EER of approximately 5%. Integration of both detection methods can achieve almost perfect classification in our current data set (644 live and 570 spoof images).

We have evaluated our methods on capacitive DC, electro-optical, and several optical scanners. Our methods show variability in performance across technologies, possibly due to differences in the quality of fingerprint images from different scanners. We also found that a live subject can protect against a false rejection by wiping his finger on clothing and then placing it on the scanner with normal pressure. This strategy can prevent perspiration-saturated and smudged (noise added) fingerprints. Fake fingers do not benefit from this technique because inherent live properties are absent. Each method of detection can achieve acceptable results, but more important, different options can be tailored for the application and hardware. The added layer of security may provide a level of confidence that end-users require for greater market acceptance.

In sum, current biometric fingerprint sensors can be fooled easily by spoof attacks. However, our work demonstrates that there is a cost-effective way to protect against spoofing. Other advantages are that new hardware is unnecessary and the algorithms can be retrofitted and tailored for existing and future biometric systems. We believe that liveness detection is a crucial technology that will promote greater acceptance of biometrics in the general population by enhancing security.

In the future, we will investigate the impact of environmental factors such as temperature and humidity on the performance of the algorithm. In addition, because the algorithm is specific to each sensor, we plan on tuning it for each manufacturer.

This work is funded by a National Science Foundation (NSF) Information Technology Research grant (0325333), the Center for Identification Technology Research at West Virginia University, and an NSF small business grant (IIP-0740601) for NexID Biometrics LLC.