Multimodal biometric authentication using adaptive decision boundaries

Recent advances in multimodal biometric authentication techniques have improved their reliability level. The use of combination methods was one of the main techniques to improve the performance. We aim to find the fusion strategy that yields the best classification rate in experiments on the MOBIO-Banca biometric data set. Scores from classifier outputs are ranked and fused using several fusion strategies. Our proposed method of combined use of adaptive boundaries and quality measure based ranking of scores yields significant improvement over existing fixed boundary methods. Sum fusion yields very large improvement and succeeds, where MProduct does not. We explain the degradation in the serial MOBIO combiner analytically and through synthetic experiments. We show that the performance of weak classifier combiners follow a hyperbolic curve as weaker classifiers are added.