On binary representations for biometric template protection

Abstract

The main objective of the thesis is to extract biometric binary strings from real-valued features, leaving independent, discriminative and reliable biometic features an important assumption in the research. Unfortunately, in practice, most of the biometric features retain large intra-class variations. Consequently, the extracted bits are still error-prone, even though the quantization and coding procedure is well-designed. If the bits are less reliable, an advanced ECC is required, otherwise the number of secrets dramatically decreases. Another weak point is that the bit extraction procedure relies on the user-dependent feature distributions. However, in most of the current systems, only a few samples are captured for every enrolled user, making it difficult to accurately estimate the feature distributions. For these reasons, the future work will focus on the following aspects: * Improving feature quality / Higher feature quality directly gives more reliable features and therefore benefits in reliable bits. The straightforward solution to improve the feature quality is to optimize the biometric capture procedure. For instance, to provide a more controlled verification environment, or to improve the quality of the features that are captured, by enhancing both the hardware and the software of the biometric sensors. An alternative solution is to introduce a quality control step. Finally, despite the common PCA or LDA feature reduction methods, it is possible to employ some other methods to extract more reliable features. * Improving the modeling of the feature distributions / Optimal quantization intervals and bit allocation are dependent on the biometric feature distributions, which are estimated from the samples of the enrolled users. However, in practice, it is impossible to obtain a sufficiently large number of samples during the enrollment. Further work would be to define, for every user, a proper number of samples required for a good estimation. As a result, the number of samples required could vary from user to user. Furthermore, an adaptive system could be gradually enrich the modeling of the feature distributions. * Designing ECC / An ECC that can recover more erroneous bits will give better recognition performance as well as more secret bits. Therefore, according to the bit error probabilities of the biometric features, an advanced ECC is desirable. * Comparing or employing multiple biometric modalities / In addition to the template protection purpose, extracting biometric binary strings enables the opportunity to compute and compare the capacities of different biometric modalities. For instance, how many secret bits does this biometric fingerprint application really have? as compared to the other application? Furthermore, for a single biometric modality, the number of secret bits is relatively low. Thus, employing multiple biometric modalities might increase the number of secret bits.