Depth-Integrated CNN Approach for Effective Face Spoof Detection

Facial recognition systems present certain benefits, notably their relative convenience, contactless nature, and nonintrusiveness compared to alternative biometric methods like voice recognition and fingerprint technology. These characteristics enhance their applicability, popularity, and overall adoption. However, the susceptibility of facial recognition to spoofing attacks poses a significant challenge. This vulnerability arises from the ease with which photos and videos of registered users can be obtained online or captured without their knowledge or consent. This article introduces a method based on convolutional neural networks (CNN) to enhance the security of facial recognition systems against presentation attacks (PA). By employing both RGB and monocular depth data, the method distinguishes between authentic and fake facial images. The initial step involves preprocessing the images to minimize background interference and improve feature extraction. Following this, the CNN analyzes the preprocessed images to extract color features from both RGB and depth data. This approach not only lowers the expenses associated with depth cameras but also improves the accuracy of classification. The efficiency of this new model has been confirmed through its training and testing on the NUAA and MSU-MFSD datasets. Quantization is implemented for model optimization to make it memory-efficient and faster to execute.