Towards an Enhanced and Lightweight Face Authentication System

Face recognition is one of the most well-adopted ways to verify someone’s identity, which however might be spoofed by presenting for example a fake image or video. Therefore, it is essential to include additional face liveness detection for a safer application. Among the existing solutions, it is common to plug in a separate model for face liveness detection. Therefore, the current authentication platform will provide two models in order to provide a safe authentication. However, in many practical scenarios, the platform (e.g., IoT devices) has limited resources in terms of computation power and storage, and this may prevent the two-models from being deployed successfully. Observed that both recognition and liveness detection work on the same face image, we believe it is possible to integrate two functions into a unified model, which will reduce the computational workload and storage requirements. To achieve this, we explore two works with different model designs, research focuses, and potential solutions. In the first work, we try to enhance a usual face recognition model with additional task capability without any additional storage cost. Concretely, we first analyze the two task’s relationship, and by a mathematics formulation, we insert the observed dual-task relationship to a novel deep model with distance-ranking feature. The training of the model focuses on the feature-learning and it does not directly use the task ground truth labels, which makes the model has a good generalization capability on new data. We have conducted experiments on a benchmark dataset and the results show that our average performance has a minimal 15% improvement compared to the baselines. In the second work, we adopt the classic multi-task learning model to combine the two tasks. Rather than using a deep multi-task model, we compress the original deep model to a lightweight version. Additionally, in order to compensate the performance degradation due to compression, a multi-teacher assisted knowledge distillation is applied where a good balance between accuracy and model size is achieved.