Abstract:
In many cases, secret key management is the weakest link in a cryptographic system.
Biometric data can help in providing a secure and user friendly way to manage crypto-
graphic keys. One way to use biometrics for secret key management is to release keys
once the associated biometric templates are found to match. However, as biometric
templates are normally stored in cleartext form, this introduces many concerns about
the security of stored biometric templates. A practical and possibly secure solution
is to bind the secret key with the biometric template in a way that eliminates the
need for direct storage of the key or the template. The binding information stored
in the system would not reveal much about either of its components. However, due
to inherently di erent characteristics of biometric templates and cryptographic keys,
a few issues must be addressed before such a system can be built. For example, bio-
metric readings are non-uniform, noisy, and irrevocable, while cryptographic keys must
be uniform, exact, and revocable. In addition, there is a tradeo between key sizes
and error correcting capability (distance between the enrollment and the veri cation
templates) in such systems which demands techniques to reduce noise and to improve
recognition accuracy.
In this thesis, we address the above-mentioned issues. Our contribution is fourfold.
First, we devise a cryptographic key management system using iris templates. Instead
of authenticating by matching templates directly, the system matches hashes of gener-
ated cryptographic keys. Authentication can be carried out without physically storing
the cryptographic keys or the biometric templates. Instead, we store, on a smart card,
some recovery information generated during the process of binding a cryptographic key
and a biometric template. For each veri cation request, the secret key is regenerated
with the help of the veri cation template and the recovery information. The system,
evaluated on the University of Bath iris image dataset, generates cryptographic keys of
260 bits with a false recognition rate of 0.24% and a false acceptance rate of 0%. This
is among the largest reported keys which have been generated using biometric readings
with reasonably low error rates.
Second, we present a biometrics-based full disk encryption scheme. A full disk en-
cryption scheme encrypts everything on the disk including the swap space and the
temporary les, and therefore does not leave any trace of plain data on the disk. To
provide security against malicious modi cation of disk contents, a full disk encryption
scheme is required to o er both privacy and authenticity of disk contents. Unfortu-
nately, existing schemes provide only privacy of data. We present a disk encryption
scheme using, as its building blocks, a robust fuzzy extractor and an authenticated
encryption scheme. The scheme provides both privacy and authenticity of data. To
support this claim, we provide new de nitions for privacy and authenticity for disk en-
cryption schemes and prove the security of our constructions by reducing their security
to the security of the underlying robust fuzzy extractor and authenticated encryption
scheme.
Third, we present a weighted majority voting scheme to improve the recognition rate
of any iris recognition system by treating it as a black box. An experimental evaluation
with CASIA version 1 iris image dataset shows that the proposed scheme improves on
the existing majority voting and reliable bit selection schemes.Finally, we present an algorithm to localize the iris in a given eye image. We use image
intensity to detect the pupil-iris boundary while edge detection and a circular Hough
transform are used to detect the iris-sclera boundary. Experiments with University of
Bath and CASIA iris image datasets show promising results in each case.