Results and Analysis
1) OHEM effectiveness evaluation : Four groups of experiments have been designed under the same experimental conditions. Table 3 presents the experimental results. The introduction of OHEM has led to an EER reduction in all models. In particular, the Resnet18- OHEM model achieved a 42% reduction in EER, and the t-DCF index decreased by nearly 50%. The best result for the single model is Res2net- OHEM, achieving an EER of 2.13%. Fig. 1 compares the performance of Resnet18 and Resnet18-OHEM models for different logical attacks. As shown in Fig.1, the Resnet18-OHEM system outperforms Resnet18 under most spoofing attacks (A07-A15, A17). Moreover, after the introduction of OHEM, the model has improved to different degrees for almost every attack identification.
2) Fusion results and comparison with other systems : Table 4 shows our scores for fusing different models and compares them with other models. After performing a two-by-two fusion of the models, it was discovered that the S3 model combined with other models could provide t-DCF values of 0.0233 and 0.0262, respectively. Analysis revealed that the S3 model outperforms at identifying A17, and this benefit may be fully realized by merging it with other models. Besides, EER=0.77 and t-DCF=0.0228 were derived by combining the results of the S1, S2, and S3 models.
Table 3: The comparison between the four models after the introduction of OHEM.