Fixed FAR correction factor of score level fusion

In biometric score level fusion, the scores are often assumed to be independent to simplify the fusion algorithm. In some cases, the “average” performance under this independence assumption is surprisingly successful, even competing with a fusion that incorporates dependence. We present two main contributions in score level fusion: (i) proposing a new method of measuring the performance of a fusion strategy at fixed FAR via Jeffreys credible interval analysis and (ii) subsequently providing a method to improve the fusion strategy under the independence assumption by taking the dependence into account via parametric copulas, which we call fixed FAR fusion. Using synthetic data, we will show that one should take the dependence into account even for scores with a low dependence level. Finally, we test our method on some public databases (FVC2002, NIST-face, and Face3D), compare it to Gaussian mixture model and linear logistic methods, which are also designed to handle dependence, and notice its significance improvement with respect to our evaluation method.