MISIS University has developed a neural network designed to verify the authenticity of facial images. At NUST MISIS, users can upload photos for verification through a dedicated web application, and there is also the option for real-time analysis via a computer camera. The project team aimed to create a robust system capable of distinguishing genuine facial data from manipulated visuals under realistic conditions, reflecting ongoing efforts to strengthen security in biometrics and digital identity verification. Attribution: MISIS University study.
In their research, the developers focused on presentation attacks. These include photographs of faces in printed form, images displayed on electronic screens, as well as three-dimensional masks crafted to resemble real features. The goal was to understand how different deceptive representations might fool a recognition system and to build defenses that reliably resist such tactics. Attribution: MISIS University study.
From among five existing neural networks, the team selected the two most promising architectures and built a two-stage system. This configuration emerged from careful experimentation and observation, aiming to balance detection speed with accuracy. The approach was designed to be practical for real-world deployment, where rapid and reliable authenticity checks are essential. Attribution: MISIS University study.
One important step in developing a machine learning solution is assembling a representative training dataset. The team curated a set of 16,500 images containing both real and fake faces, with roughly equal representation of deception types used in facial recognition systems. The dataset included photos shown on electronic device screens, realistic masks, and renderings that imitate human features. Additional fake images were created by altering photos of individuals with various characteristics to further diversify the training material. This careful data curation helps the model recognize subtle cues that differentiate genuine faces from spoofed representations. Attribution: MISIS University study.
In the first stage, the system employs a pre-trained MTCNN network to locate the face within an image. After identifying the facial region, an expanded area is introduced where 60 percent of the new region focuses on the analyzed face. This enlargement improves the model’s ability to extract reliable facial patterns. The InceptionResnet component is then used to generate numerical representations of facial features, providing a compact and informative fingerprint of each face. In the second stage, additional network layers process these features to capture more nuanced patterns that signal authenticity. Attribution: MISIS University study.
Results from the two-stage process are combined and routed through several final layers to render a verdict on the image’s authenticity. This layered approach produced a high level of accuracy in identifying genuine faces and detecting spoofed inputs, demonstrating the potential of layered deep learning architectures for biometric security tasks. Attribution: MISIS University study.
Earlier work in this domain included training neural networks for interviewing sales managers, illustrating the broader applicability of advanced pattern recognition in diverse fields and highlighting the adaptability of these techniques to real-world, task-specific data. Attribution: MISIS University study.