*3.2. Experiment Setup*

To evaluate the proposed approach, we conducted several experiments. First, we carried out an initial experiment to evaluate the performance of the fingerprint branch net regarding the detection of PAs. We compared our results with previous state-of-the-art methods. For this purpose, we utilized the LivDet 2015 dataset, whereas the fingerprint branch net was trained on the training portion of the LivDet 2015 and tested on the testing portion of the same dataset. In the second experiment, we evaluated the three proposed fusion architectures in detecting and preventing the PAs. We then conducted an experiment to analyze the sensitivity of the highest performing architecture during the second experiment. To this end, we analyzed the effects of increasing the number of subjects during the training on the classification accuracy. Finally, we reported the number of parameters and classification time by the proposed architectures compared with state-of-the-art methods.

For training the network, we use the RMSProp optimizer with the following parameters: β is set to its default value (β = 0.9), whereas α is initially set to 0.0001 and is periodically decreased by a factor of 1/10 for every 20 iterations (epochs). For compatibility with the LivDet 2015 competition [27], the accuracy was used as the evaluation parameter in all of the experiments. The accuracy is defined as the percentage of correctly classified samples.

All experiments were repeated five times and the average classification accuracy was reported. The experiments were carried out using a workstation with i9 CPU @ 2.9 GHz, 32 GB of RAM, and NVIDIA GeForce GTX 1080 Ti (11 GB GDDR5X).
