D3.2: A study on PKI and biometrics

Title: SOFT BIOMETRICS

As already mentioned, biometrics refers to the automatic identification or verification of individuals by using their physical or behavioural characteristics. Nevertheless, up to now no recognition or verification method based on one physical or behavioural feature combines efficiency, universality, robustness to noise in data collection, reasonable protection against spoof attacks, acceptable error rates, satisfying matching rates, and satisfying temporal (especially in real-time applications) and spatial complexity both for training and verification, user-friendliness. This limited ability of the implemented methods is due to not only the limitations of the methods but also to the theoretical upper bound that every biometric has with respect to its ability to distinguish two individuals. 

A solution to these limitations is the use of soft biometrics features. Soft biometrics are the human characteristics that provide information about the individual which however is unable to sufficiently differentiate any two individuals and thus identify an individual reliably and uniquely due to its lack in distinctiveness and permanence [JAI04]. These biometrics include information such as gender, eye colour, ethnicity, age and height. In order to perform identification or verification, a system totally based on soft biometrics information cannot provide results with a satisfactory matching rate. Nevertheless, soft biometrics can offer great amelioration to the performance of common biometric systems (e.g., signature verification system, face recognition system, fingerprint identification system) by complementing the identity information of each individual. 

Soft biometrics are able to further improve the performance of a biometrics system by tuning the parameters of the system, such as thresholds and even weighting in case of multimodal biometrics systems. A significant improvement that soft biometrics offer is the filtering of large biometric databases; in other words, limiting of the database entries which the traditional biometrics system (e.g. face recognition system) will include to its effort to recognise an individual and thus improvement of the efficiency and speed of the system is achieved, whereas the possibility of false positive is reduced. For instance, if the person to be identified is a blue-eyed child, then the search can be limited to the records that concern persons with this colour of eyes and age group.  

Technical description of the methods 

The use of soft biometrics requires them to be properly obtained from the user. More specifically, the system should be able automatically to determine the required soft biometric traits (such as height, age and gender) of the person when they interact with the main biometric system (for example the fingerprint identification system). The height of person can be estimated for instance through the use of ultrasonic sensors. 

Xiaoguang Lu and Anil K. Jain [LU04] proposed a method for ethnicity recognition (Asian vs. non-Asian) taking into account the fact that human face images include valuable information concerning the person’s gender and ethnicity. The problem is addressed in a machine learning framework, whereas Linear Discriminant Analysis based scheme is used for the two classification task. This latter method is a popular statistical method used for the projection of given multidimensional data to a lower dimension aiming at maximising the ratio of between-class scatter to within-class scatter. Given the fact that images on different scales provide different levels of information, they resize each face image to three different scales with an LDA based classifier built for each one. The nearest neighbour rule is used during comparison, while the matching score is composed by the cosine value of the angle between the test image vector and the training image vector. As expected, the classification result of the test input image is determined by the training image with the best match.

Value and technical limitations

The accuracy of soft biometric systems is vital when they are used complementary to primary biometrics since a false negative result can be caused by false filtering of the enrolled biometrics data leading to overall system performance degradation. The presented ethnicity identification method offers rather promising results with its average accuracy rate reaching 93% involving four categories (Caucasian, South Asian, East Asian and African) and over 3000 test images. 

Privacy aspects 

One of the main advantages of soft biometrics is the fact that they lack uniqueness and thus the knowledge of a person’s soft biometrics is not enough to threaten the person’s privacy. They are used in a manner complementary to traditional biometrics systems and most of them are exposed during a person’s daily activities.