A Novel Deep Ensemble Learning to Enhance User Authentication in Autonomous Vehicles

For autonomous vehicles, face-based authentication is one of the best ways to identify or verify a user. Deep learning methods, such as Deep Convolutional Neural Networks (DCNN), have brought excellent accuracy in recent years. However, building an optimal DCNN and ameliorating face-based authentication performance are still the main challenges for autonomous vehicles. From these challenges, we first investigate and analyze the effect of several novel effective loss functions on DCNN, such as CosFace, ArcFace, and MagFace, for face-based authentication. We then propose a parallel ensemble learning, namely PelFace, by taking advantage of recent models based on CosFace, ArcFace, and MagFace. PelFace elaborates on the voting mechanism that utilizes non-optimal pre-trained models to obtain better discriminative ability for face-based authentication. In addition, we propose to perform the matching phase for each PelFace model in parallel. To prove the performance of both the speed and accuracy of PelFace, we carry out rigorous experiments using several popular benchmarks, including LFW, CFP-FP, and AgeDB-30, as well as two renovations of LFW: CALFW and CPLFW. The results of our experiments achieve state-of-the-art figures, which show the proposed method’s massive potential in improving face-based authentication performance. Ablation studies and overall benchmarks significantly confirm the effectiveness of our PelFace. Note to Practitioners—This paper motivates to construct and improve the discriminative power of face-based authentication based on Deep Convolutional Neural Networks (DCNN) for autonomous vehicles. Facial authentication is widely applied to identify or verify a user, but its accuracy and speeds still need improvement, especially for autonomous vehicles. Thus, in this work, we propose a parallel ensemble method called PelFace, by combing several novel DCNN models, such as CosFace, ArcFace, and MagFace. Our method is built by using the original Torch library for Python language. We also describe how such trained models can be combined to obtain SOTA performance, allowing us to reduce the time for loss function hyper-parameters tuning to get optimal models. In addition, we also indicate the drawback of the combination method as well as an approach to solving. Our experiments indicate that using this technique to create high-accuracy facial authentication is possible without expensive hardware. Current results of PelFace allow proving its ability to devote user authentication in autonomous vehicles, such as passenger identification and driver verification.