Deepfake detection is of fundamental importance to preserve the reliability of multimedia communications. Modern deepfake detection systems are often specialized on one or more types of manipulation but are not able to generalize. On the other hand, when properly designed, ensemble learning and fusion techniques can reduce this issue. In this paper, we exploit the complementarity of different individual classifiers and evaluate which fusion rules are best suited to increase the generalization capacity of modern deepfake detection systems. We also give some insights to designers for selecting the most appropriate approach.