Nowadays biometric technologies are becoming widely used in application associated with security problems [1-2]. Recently number of tasks have appeared where no single biometrics (like face, finger etc) meets very tough requirements to identification quality. The most important areas with high requirements to biometric solutions are criminal ID, biometric passport, access to classified information.
Generally if we any biometric applications, the most widespread approach to fusion is training on some set. As a result, we have discriminant function, that separates genuine matches (if both matched biometric samples belong to one person) and impostor matches (samples does not belong to one person).
In that case, on the one hand we can totally control fusion process. On the other hand we face two apparent problems. First, if parameters of the training set differ from operation data, we have to retrain manually. Second, multimodal biometric sets are so small that they are not enough for such simple training at the stage of designing.
The basic idea of automatic training process is tuning of “fusion process” without operator or researcher. Such approach allows to decrease expenses on modernization. The second significant advantage is possibility of retraining during operations.
The significant methodology problem of automatic training is character of available information affiliated with the fused biometric performance. The lack of information is a typical problem of direct applying well developed statistical technique to multimodal biometrics training. In particular, one the most trouble prone specifics is lack of multimodal bases.
We analyzed the experience of multimodal biometrics designing. Basing on the analyses results, we settle requirements an arbitrary the training procedure:
• optimal performance as measured by FAR and FRR;
• functional requirements:
o parametric decision rules;
o taking into account biometric specific;
o training without multimodal samples;
• implementation requirements:
o robustness to small training sets;
o generalization ability.
We studied statistical features of biometric technologies to fill the requirements. We revealed character of dependencies between different biometrics. Approach to automatic training is proposed. The keystones of proposed approach are short-listed below:
• division of training process and affiliated problems into several fusion levels;
• study of specifics of each fusion level;
• taking into account biometric specifics:
o stochastic target quality indices;
o stochastic character of identification process;
o unavailability of multimodal datasets