Iterative Binary Training

Vítor Luiz Da Silva, Rosana Tomàs, Fernando Cores, Francesc Giné; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 3239-3247

In this work, we present Iterative Binary Training, an effective training strategy designed to improve face anti-spoofing systems, especially when dealing with imbalanced datasets. Instead of treating all spoofing attacks at once, our method starts by training a binary classifier to distinguish bonafide faces from the most frequent spoofing type. Then, step by step, it adds new spoofing classes--moving from the most common to the rarest-- and thus, the model gradually learns to handle a wider variety of attacks. This method encourages the model to first focus on dominant spoofing patterns and later adapt to more challenging, less frequent attacks, reducing overfitting and improving generalization. We tested it using four deep learning models, including ViT-B/16, ViT-B/32, ResNeXt-101, and ResNet-50. All showed good performance with our method, with ResNeXt-101 standing out as the top performer. Our approach does not rely on extra data, additional modalities, or ensembling techniques. Instead, it builds on standard tools like class-balanced loss functions and pretrained backbones, making it easy to reproduce and deploy. The results suggest that Iterative Binary Training offers a promising direction for enhancing FAS systems in real-world scenarios.