4. Conclusion

into two depending on their characteristics: physiological characteristics and behavioral characteristics. Biometrics with physiological characteristics contains direct physical evidence of discriminative features in their samples. This category includes biometrics such as face, fingerprints, palm veins, DNA, retina, iris, and ears. On the other hand, discriminative features from biometrics with behavioral characteristics can be indirectly extracted from samples, and this

A biometric system for identification or recognition can be designed and implemented either feature-based using a handcrafted feature extraction or automatic feature generation-based using an end-to-end training based on a machine learning algorithm. For a feature-based biometric system, the selection of feature types and descriptors, and a following classifier design become very important to reduce the variability and the computational complexity of original characteristics. Each feature descriptor has its own strength for specific type of patterns. For example, the Gabor filter has better direction selectivity and frequency selectivity, so it can be used to apply time-frequency analysis for input images. Texture coding operators such as LBP and its variants are generally robust to changes from illumination and facial expression in images. Hence, it is critical to select right features according to the applications. Performance of a feature-based system mainly relies on capability of human experts, and it often results in low generalization for variations on input data. Recent automatic feature generation-based approaches such as deep learning can be an excellent alternative to deal with such difficulties. In this type of system, feature extraction/selection and classifier parts are trained together with large amount of data, and it generally shows better performance than a feature-based system. One disadvantage of this approach is that it usually consists of large

Biometric systems for security require to have very high accuracy with favorably low computational complexity. In addition, a reliable biometric system should generalize well for unseen samples, and highly robust to various type of challenges including geometric transformation, illumination change, intraclass variation, and presence of noise. For a real-world security application, we need to construct a system endurable to various types of attacks such as counterfeiting. In order to make a system with higher security, multimodal biometrics [2] has attracted wide attentions in recent years and becomes a hot research topic. A multimodal system, for example, with input of finger vein and finger print images, has higher reliability, broader applicability, and stronger security and can provide a more reliable and stronger

Machine learning is a procedure to learn from examples and, more specifically, it is a field of optimizing system parameters, which are defined on an architecture, to meet the evaluation criteria using a set of training examples. We often use statistical techniques to give computers the ability to "learn." Once the intended goal of learning is met, we may use the resulting

category includes biometrics such as typing rhythm, gait, and voice.

number of parameters and takes a rather long time to train them.

security in practical applications than unimodal one.

3. Machine learning

4 Machine Learning and Biometrics

A biometric system for security should be very reliable and accurate. Feature-based biometric systems can be designed and implemented with their relatively high accuracy and fast response. For more reliable and accurate systems, machine learning techniques can be applied to biometrics and their application areas. Especially, novel powerful algorithms, such as deep learning algorithms, can be excellent candidates for solving the challenging biometrics problems.
