Development of Active Liveness Detection System Based on Deep Learning ActivenessNet to Overcome Face Spoofing
Dhimas Adi Nur Fauzi, Hendrawan Hendrawan, Eueung Mulyana, Wawan Hermawan · 2023
The facial verification system in the e-government (SPBE) is vulnerable to face spoofing attacks. Face spoofing can be overcome through the liveness detection method. One type of face spoofing that the writer will tackle is an attack in the form of a 3D mask attack. In this case, the active liveness detection method is a process for detecting liveness by giving a series of questions in random order. Then the user must follow all these questions so that they can be categorized as real. If the user cannot follow the questions, it will be categorized as fake. The model used for liveness detection is ActivenessNet, which is a combination of three pre-trained models, namely emotion detection, profile detection, and blink detection. Tests carried out on these pre-trained models produce an accuracy value in the validation set of 70% in the emotion detection model. This value is the highest value when compared with other models. For the profile detection model, an accuracy value of 95% is obtained in the validation set, which contains faces facing left and right. The blink detection model was tested to detect ten blinks on the video and obtained a 100% recall value in bright and dark lighting conditions when the user is not wearing glasses. When the user wears glasses, there is a decrease in model performance, with a recall value of 80% in bright conditions and 60% in dark light conditions.