| dc.description.abstractEN | We have entered an era of misinformation, fake news and impersonation fuelled by Ar- tificial Intelligence (AI). Visual content have been jeopardized by malicious entities in an attempt to fool security systems by pretending someone else’s identity. Such entities, usually seek elevated access to critical infrastructures like online banking, government administration services and any KYC system. A Biometric’s system task, is to deploy security checks and measures to verify someones identity and authenticity(liveness). Given the societal impact of such commodification of impersonation, our research pro- poses learning-based methods with focus on learning to perform well on a significantly different target distributions a.k.a Domain Adaptation and Domain Generalization. In Deep Learning (DL), model performance depends heavily on the presence of high vari- ation within training examples and usually, Machine Learning practitioners aim to col- lect data in such way that will guide model’s ability to generalize. However, even with enormous amount of data there is no guarantee that a model would perform equally well to unseen data in the same domain.
In our research, we explore and evaluate different deep and machine learning methods in an attempt to learn discriminative features that could generalize to several academic datasets and datasets in-the-wild. We define a classic binary classification problem which is used as a baseline model. We use Deep Convolutional Neural Networks(CNN) and two commonly used backbones, Resnet18 and EfficientNetb4. Then, we transform the task into a multi-task learning with auxiliary fully-connected heads and explore the impact in generalization and adaptation, and use different deep metric losses such Cen- ter Loss and Triplet Loss and Gradients Orthogonality. Finally, we investigate hand- crafted features such Histogram of Oriented Gradients(HOG) and Local Binary Pat- terns(LBP) and train classic machine learning classifier (eg. Support Vector Machine). | el |
