MISIS University has developed a neural network that assesses the authenticity of facial images, enabling verification for uploaded pictures through a purpose-built web application hosted by NUST MISIS. This platform allows users to submit photographs for analysis and offers real-time verification via a computer camera connected to the system. The goal is to distinguish genuine facial data from manipulated inputs with high precision, a critical capability for security applications and identity verification services used in Canada and the United States.
The research team explored a range of presentation attacks, including photographs printed on paper and displayed on electronic screens, as well as three-dimensional masks designed to resemble real faces. By examining these variants, the developers aimed to build a model robust enough to detect both classic image-based fakes and more sophisticated physical forgeries. This focus on deception types is essential in environments where fraudsters continually refine their methods, and it helps ensure that the verification system holds up under diverse, real-world conditions.
From an initial pool of five established neural networks, the team identified two promising architectures that showed complementary strengths. They then assembled a two-stage system that integrates insights from both networks. This configuration reflects a careful balance between speed and accuracy, enabling rapid preliminary assessments while retaining the capacity for deeper, feature-focused analysis when needed. The resulting architecture is designed to be both scalable and adaptable to evolving attack vectors in biometric authentication.
Training data plays a pivotal role in the system’s reliability. The team assembled a dataset containing 16,500 images that encompass both authentic and counterfeit faces. The dataset includes a roughly even distribution of deception types, featuring printed photographs, screen-based displays, masks, and even synthetic, cartoon-like representations. A key element in building this resource was generating a variety of fake images that capture different external appearances, ensuring the model learns to recognize subtle cues of manipulation across a wide spectrum of facial presentations. One of the developers, Alisa Semenova, explained that the dataset also incorporated images printed with diverse features to mimic real-world diversity, thereby strengthening the model’s ability to generalize beyond the training set.
In the first stage of the face verification pipeline, a pre-trained MTCNN network identifies the location of a face within the input image. To sharpen the focus on the relevant region, a dedicated area is then added to the image, occupying about 60 percent of the scene and centered on the detected face. This targeted emphasis helps boost accuracy by reducing distractions from non-facial elements. The next step leverages the Inception-ResNet architecture to convert facial features into compact numerical representations, capturing the distinctive geometry and texture patterns that differentiate genuine faces from forgeries.
During the second stage, additional neural network layers analyze those representations to extract a broader set of discriminative features. The combination of the two stages feeds into a final decision process, where several concluding layers synthesize the collected evidence to determine image authenticity. This multi-layered approach enhances resilience against a wide range of manipulation techniques, and the workflow has demonstrated strong performance in testing scenarios. The two-stage fusion enables the system to balance the speed of initial screening with the depth of subsequent verification, which is particularly valuable for high-volume environments where rapid yet reliable assessments are required.
Overall, the two-stage framework achieves a high level of accuracy in authenticating facial data, reflecting a careful integration of proven neural network components with problem-specific adjustments. The architecture is designed to be adaptable, so it can be updated as new attack vectors emerge or as computational resources evolve. The researchers emphasize that ongoing refinement, expanded datasets, and real-world validation are essential to maintaining robust performance in face verification tasks across different settings and user populations.
In related work, earlier studies explored neural network approaches for interviewing and evaluating sales managers, highlighting the breadth of biometric verification research and its potential applications in customer interactions and security protocols. This broader context helps illustrate how advances in neural networks for face verification intersect with real-world practices in security, human-computer interaction, and enterprise risk management, especially in North American markets where identity verification is a critical component of digital trust and service integrity.