Cancelable approaches for biometric person authentication have been studied to protect enrolled biometric data, and several algorithms have been proposed. One drawback of cancelable approaches is that the performance is inferior to that of non-cancelable approaches. As one solution, we proposed a scheme to enhance the performance of a cancelable approach for online signature verification by combining scores calculated from two transformed datasets generated using two keys. Generally, the same verification algorithm is used for transformed data as for raw (non-transformed) data in cancelable approaches, and, in our previous work, a verification system developed for a non-transformed dataset was used to calculate the scores from transformed data. In this paper, we modify the verification system by using transformed data for training. Several experiments were performed by using public databases, and the experimental results show that the modification of the verification system improved the performances. Our cancelable system combines two scores to make a decision. Several fusion strategies are also considered, and the experimental results are reported here.