TEMPLATE CONSERVATION METHODOLOGIES FOR MULTIMODAL BIOMETRIC WITH LSTM NEURAL NETWORK

In this work, we present a privacy-preserving multi-modal biometric system that uses LSTM (Long Short- Term Memory) neural networks for classification. Feature-level fusion has been applied over the features extracted by computer vision algorithms such as SURF (Speeded up Robust Features) and HoG (Histogram of Oriented Gradients). This work proposes two template preservation methods, bio-hash, and simple hash,
to develop a secure architecture. The cancelability of the templates can be achieved by modifying the seed value of the random matrix. In the experimentation, various feature counts and their performances are tabulated under various metrics. The results show that when HoG feature method is applied, bio-hash method gives the lowest EER (Equal Error Rate) as 0.45 at feature count 10. And the simple hashing
method's lowest EER is 0.34 at the feature count 10. When SURF feature method is applied, bio-hash method gives EER as 0.58 at feature count 30. And for simple hashing method, the lowest EER is 0.4 at the feature count 40.