With a significant difference in recognition rate the sum rule becomes a weighted sum rule.

So the normal sum rule may be like: SumConf = (Conf1 + Conf2) /2

Then a weighted sum rule can be like: WeightedSumConf = (Weight1 * Conf1 + Weight2 * Conf2), with Weight1 + Weight2 = 1.

You should adjust the weights such that you get optimal performance; optimizing to the best performance is a way to obtain that. Alternatively, you can relate the weights to the performance of the individual modalities.

Thanks Kalamer i will try ti weighted sum rule  but the values of wheights i will adjust them randomly   ??   

Most likely the better modality has higher weights.

Have you normalized both of the scores? The weights should not be randomly adjusted. Instead, you should run experiment to estimate the appropriate weights.

this paper uses the summation rule:

https://www.researchgate.net/publication/265168466_Solving_the_Problem_of_the_K_Parameter_in_the_KNN_Classifier_Using_an_Ensemble_Learning_Approach?ev=prf_pub

Article Solving the Problem of the K Parameter in the KNN Classifier...

http://www.sas.ewi.utwente.nl/open/courses/intro_biometrics/Ross01.pdf

Have a look at "Formulae for Consistent Biometric Score Level Fusion" by Hube.