Combining Dynamic Image and Prediction Ensemble for Cross-Domain Face Anti-Spoofing

Abstract

Most of the face anti-spoofing methods improve the generalization capability by adversarial domain adaptation via training the source and target domain data jointly. However, considering the data privacy, it is impractical in application. Hence, we propose a source data-free domain adaptative face anti-spoofing framework to optimize the network in the target domain without using labeled source data via modeling it into a problem of learning with noisy labels. To obtain more reliable pseudo labels, we propose dynamic images with the background to capture the motion divergences between real and attack faces. Nonetheless, fluctuations of predictions caused by noisy labels are still strong. Therefore, a filtering strategy is proposed to reduce the impact of noisy labels by self-ensemble, which combines prototype and progressive pseudo labels predicted by the source pre-trained model and target model respectively. The proposed approach shows promising generalization capability in several public-domains face anti-spoofing databases.

Algorithm of face anti-spoofing based on pseudo-negative features generation

“…One such approach involves stylizing target data to match the source domain style using image translation techniques and then classifying the stylized data using a well-trained source model (Zhou et al, 2022a). Additionally, novel frameworks such as Sourcefree Domain Adaptation for Face Anti-Spoofing (SDAFAS; Liu et al, 2022a) and a source data-free domain adaptive face anti-spoofing framework (Lv et al, 2021) have been proposed to tackle issues related to source knowledge adaptation and target data exploration in a source-free setting. These frameworks aim to optimize the network in the target domain without relying on labeled source data by treating it as a problem of learning with noisy labels.…”

Deep Learning for Face Anti-Spoofing: A Survey

“…In the deployment phase, the source data cannot be shared for adapting the pre-trained model to the target data, as they contain sensitive biometric information. Lv et al [201] benchmark the source-free setting for FAS via directly applying a self-training approach, which easily obtains noisy target pseudo labels due to the challenges in the FAS task (e.g., the intra-class distance between live faces of different identities probably exceeds the inter-class distance between live and spoof faces of the same identity). Thus, the performance gain (1.9% HTER reduction on average) by adaptation is quite limited.…”

Online Adaptive Personalization for Face Anti-spoofing

“…Quan et al [23] propose temporal smoothing of predictions and a progressive pseudo-labeling approach where the thresholds are varied over time. Lv et al [24] use predictions of model ensembles from different test epochs. However, both of these methods [23,24] assume that all test data is received at once and that the model can be trained using this data before finally making its prediction.…”

“…Lv et al [24] use predictions of model ensembles from different test epochs. However, both of these methods [23,24] assume that all test data is received at once and that the model can be trained using this data before finally making its prediction. This is different from the real-world setting where test data arrives in an online streaming fashion and the system requires low-latency prediction for every incoming frame.…”