ESR 4: Multibiometrics architectures and privacy in a mobile environment
What was achieved and the impact
Here four different identification approaches have been probed by using the finger vein templates in matching steps: i) initially the-state-of-the art CNN models were compared with the proposed model by training them via still vein images. However, as using only CNNs with still images does not provide a way for exploiting the finger vein template sequences of moving hands, we have also applied to ii) majority voting (MV) on the separate identification decisions given for each element of an image sequence, iii) score-level fusion (SF) on the decision probabilities calculated for each element of a sequence, and iv) Long-Short Term Memory (LSTM) networks which are trained by the sequential features extracted via CNN models
A novel zero-leakage biometric crypto-system that generates indistinguishable Auxiliary Data (AD) has been proposed. There, if a couple of AD sets are generated through different Pseudonymous Identifiers (PI), an attacker cannot distinguish if the two AD are linked to a same biometric trait or not. The features that feeds the crypto-system has been generated through a custom CNN network. The good stability and discriminability of the learned features make possible to embed up to 256 bits PI. The main advantage of our system: it does not need any secret key or parameter during authentication, but only a biometric sample. The system stores only the AD, that may be considered a public information, and a PI’s hash value.