DETECTION OF PRINTED FALSE ATTACKS USING NEURAL NETWORKS

115 “Al-Farg‘oniy avlodlari” elektron ilmiy jurnali ISSN 2181-4252. Tom: 1 | Son: 3 | 2025-yil "Descendants of Al-Farghani" electronic scientific journal. ISSN 2181-4252. Vol: 1 | Iss: 3 | 2025 year Электронный научный журнал "Потомки АльФаргани" ISSN 2181-4252. Том: 1 | Выпуск: 3 | 2025 год https://al-fargoniy.uz/ DETECTION OF PRINTED FALSE ATTACKS USING NEURAL NETWORKS Abdukadirov Bakhtiyor, Fergana State University Docent, Department of Information Technology e-mail: [email protected] Abstract. This article discusses a method for detecting false positives against a biometric facial recognition system based on deep convolutional neural networks. The proposed method is designed to detect printed false positives and is tested on open databases of real and fake faces, and the results are analyzed. The types of false positive attacks launched against a biometric system based on existing faces are analyzed. Keywords: biometric system, false alarm attack, local binary pattern, support vector machine, convolutional neural networks, recurrent network, classification evaluation metrics. Introduction. With the widespread adoption and increased effectiveness of biometric recognition systems, the number of attempts by potential attackers to directly hack the system, for example, by logging in as a regular user, has increased. Such attempts are called spoofing attacks [1]. Most existing biometric systems are vulnerable to spoofing attacks. For biometric systems, a spoofing attack involves deceiving the system by presenting photographs, videos, or pre-recorded sounds to the biometric sensor, with the goal of impersonating an unregistered user during the recognition process. [2] successfully demonstrated hacking commercial facial recognition systems using a photograph or video recording of a registered user using the device's screen. According to the developers, the best results were demonstrated by systems using specialized scanners or video cameras that allow 3D object reconstruction. However, methods for confirming the authenticity of a recognized object that do not require specialized equipment and do not require additional actions are the most promising, as they are more convenient for the end user and can be easily integrated into existing facial recognition systems. An additional challenge for developers of antispoofing systems is the lack of public databases containing a complete list of spoofing attacks (paper photographs of the object, photos and videos from various device screens, photo masks, silicone masks, images of the object with applied makeup, as well as 3D masks and dummies). Many researchers test developed object authentication solutions on their own databases, which are not publicly available. The lack of publicly available databases hinders a fair evaluation and comparison of proposed methods for confirming the authenticity of a recognized object. Literature analysis and methodology. There are a number of studies that use various textural features to detect spoofing: LBP and support vector machines (the percentage of correct recognition of a real object was at least 91.2%, while the percentage of attackers missed was no more than 0.2%) [3, 4], LBP and artificial neural networks (the percentage of correct recognition was 97.5% based on two-dimensional images [5], Gabor wavelets [4], histograms of oriented gradients [6], etc. In [7], images of the user's pupils are used to confirm the authenticity of the recognized object. In this invention, anti-spoofing protection is based on the property of the human pupil to constrict with increasing light intensity. To determine whether the user's image was obtained from the eye of a living person and not from a dummy, the illumination intensity is varied. The anti-spoofing system monitors the pupil's response to light modulation, therefore slightly increasing the registration time. In [8], the user is asked to look at randomly assigned locations on the monitor. Eye movement trajectories are then analyzed. The method's developers report a 95% correct recognition rate for substitutions 116 “Al-Farg‘oniy avlodlari” elektron ilmiy jurnali ISSN 2181-4252. Tom: 1 | Son: 3 | 2025-yil "Descendants of Al-Farghani" electronic scientific journal. ISSN 2181-4252. Vol: 1 | Iss: 3 | 2025 year Электронный научный журнал "Потомки АльФаргани" ISSN 2181-4252. Том: 1 | Выпуск: 3 | 2025 год https://al-fargoniy.uz/ in a database containing two-dimensional images of the object being recognized. A combination of different approaches is also possible to achieve higher results. In [9], the user is asked to make a movement specified by the system, change facial expressions, open their mouth, etc. The sequence of images from an IR camera is then analyzed after normalization and noise removal. A combination of several approaches is also presented in [10]. The presented object substitution detection system uses a video camera, motion sensors, and light sensors. The researchers obtained the following results: recognition speed was 3 seconds, the percentage of correct recognition of the real object was 95-97%, and the percentage of intruders missed was 2-3%. The disadvantages of such methods include the need for additional equipment or inconvenience for the end user (an unfriendly interface of the authentication system). Currently, there are no technologies capable of providing reliable recognition without specialized equipment. To ensure reliable recognition, it is necessary to ensure a high probability of correct facial image recognition, confirm the authenticity of the recognized object, and enable the recognition system to operate in real time. Existing technologies can only perform these operations individually. Therefore, it has become necessary to develop appropriate technology and individual components of the recognition process. Verification systems are primarily focused on recognizing the user's face and do not protect against substitution of the recognized object. In [2], the authors demonstrated a successful hack of commercial facial recognition systems using a photograph or video recording of a registered user. Biometric recognition systems require the implementation of effective methods for confirming the authenticity of the recognized object. Attack Model for Biometric Systems. When developing security methods for biometric user recognition systems, it is necessary to identify all possible threats and describe the attack model. Attack models for various biometric systems have already been developed [11, 12]. This section describes an attack model for facial recognition systems. Nine of the most vulnerable areas for attack by attackers are identified in the general biometric system diagram shown in Figure 1. Fig. 1. General diagram of a biometric user recognition system with attacks identified To carry out successful attacks on biometric recognition systems, an attacker must have skills in various specialized fields, as well as have information about flaws in the equipment and hardware implementation, the structure and method of organizing the database, methods of calculating and comparing informative features, about the subsystem for interacting with the database, and about other models and methods embedded in the implemented biometric recognition system [11]. Identifying false attacks. Unlike systems that require additional equipment (fingerprint recognition, etc.), recognizing a person from a facial image makes it very easy to create a replica of the target object. All that is needed is a photograph of the person, which can be easily found online or photographed remotely. One of the objectives of the study is to detect the substitution of an object presented to a video camera. This task is not always easy, even for humans. The solution must have low computational complexity and a high probability of correctly detecting the substitution of a recognized object without the use of additional specialized equipment. Depending on the focus of the biometric system, spoofing attacks can have varying levels of complexity. In recent years, deep learning methods have been actively used to solve this problem. In [20], an AlexNet-type CNN was used for binary classification 117 “Al-Farg‘oniy avlodlari” elektron ilmiy jurnali ISSN 2181-4252. Tom: 1 | Son: 3 | 2025-yil "Descendants of Al-Farghani" electronic scientific journal. ISSN 2181-4252. Vol: 1 | Iss: 3 | 2025 year Электронный научный журнал "Потомки АльФаргани" ISSN 2181-4252. Том: 1 | Выпуск: 3 | 2025 год https://al-fargoniy.uz/ of a real face versus a spoofing attack, and the effect of the size of the context captured during facial image alignment on the quality of this classification was also investigated. The authors of [9] used SVM to classify high-level features extracted from the final layers of a retrained VGG-Face network. In [19], a combination of a CNN and an LSTM-type recurrent network was used to classify a frame sequence. Despite significant progress in facial vitality detection, this problem remains unsolved in general. The ultra-high pixel density and natural color rendition of modern displays make the displayed facial image almost indistinguishable from the real one. Therefore, there is a need to improve the proposed spoofing detection algorithms and develop new approaches. Currently, deep convolutional neural networks are a standard building block in virtually all image processing tasks. Modern machine learning libraries significantly accelerate the development of new neural network architectures, which leads to a gradual increase in the complexity of their computational graph. Recent work in the field of deep learning has demonstrated the effectiveness of the attention mechanism in improving the performance of CNNs in tasks such as pattern recognition, image caption generation, and others [14]. Convolutional neural networks are typically trained using the RGB channels of an image directly. However, numerous previous studies on facial vitality detection have demonstrated the effectiveness of using various texture descriptors to encode the facial region. The need to generalize differences in illumination and facial pose across two visually similar classes, while simultaneously detecting fine-grained texture differences, complicates the task of neural network optimization for this problem. Research results. To detect fake attacks on biometric systems, we train a deep neural network to create a liveness detector capable of distinguishing between real and fake faces. We consider face liveness detection as a binary classification problem. We train the neural network on the NUAA Imposter Database and LCC FASD databases, which contain both real and fake face images. The NUAA Imposter Database contains 5,105 real and 7,509 fake face images, while the LCC FASD database contains 7,047 real and 7,076 fake face images. To expand the training set to protect against fake attacks, a series of augmentation processes are used to simulate various effects, including: image rotation by a specific angle, image enlargement, image dragging along the width and height, and image rotation around the horizontal axis. The training and test sets for all databases are split 75% to 25%. The accuracy and error graphs of the counterfeit attack detection model after training on the above database are shown in Figures 2–3. Fig. 2. Accuracy and loss on the NUAA PI DB database Fi g. 3. Accuracy and loss on the LCC FASD database The difference between train loss and validation loss is that the former refers to the training set, and the latter refers to the test set. Thus, the validation loss shows how the model performs on extraneous data 118 “Al-Farg‘oniy avlodlari” elektron ilmiy jurnali ISSN 2181-4252. Tom: 1 | Son: 3 | 2025-yil "Descendants of Al-Farghani" electronic scientific journal. ISSN 2181-4252. Vol: 1 | Iss: 3 | 2025 year Электронный научный журнал "Потомки АльФаргани" ISSN 2181-4252. Том: 1 | Выпуск: 3 | 2025 год https://al-fargoniy.uz/ (i.e., data not trained on this data), and therefore serves as a key indicator for evaluating the model. The validation error is also used to prevent overtraining. The accuracy, precision, recall, and f1-score evaluation metrics were used to classify the data. The results of comparing the performance of the three types of databases with the evaluation are presented in Table 1. Table 1. Performance comparison results Name of the database Performance evaluation precision, % recall, % Accuracy, % f1, % NUAA PI DB 100 100 100 100 LCC FASD 95 94 94 94 The results presented in Table 1 show that when the neural network model was trained on the NUAA PI DB database, overfitting occurred, in which the model simply memorized some of the data from the training sample. This can also be seen from the accuracy and error graph presented in Figure 2. Although the results on other databases are better than the first case, their results cannot be called perfect. Conclusion. In this article, the primary type of deceptive attack on biometric identification systems is the fingerprint attack. This attack was detected by training a convolutional neural network on two databases of real and fake faces. Of the deceptive attacks listed above, the most common are fingerprintbased and video-based attacks. The resulting live face detection detector, capable of distinguishing between real and fake faces, can detect not only fingerprintbased attacks but also video-based attacks, but performs poorly in detecting disguised attacks using 3D faces. The database also plays an important role in the development of the fake face detection detector. The aforementioned databases, due to their open source availability, allow for extensive experiments. However, modern and proprietary databases consisting of real and fake face images, such as IDIAP ReplayAttack, PHOTO-ATTACK, CASIA FASD, MSU Mobile Spoofing Database, Print Attack, Replay Attack, Gated Recurrent Unit, and OULU, were not used. The databases used in the experiments lack ethnic diversity, and multimodal methods should be used to improve the effectiveness of fake face detectors. REFERENCES 1. Фозилов Ш.Х., Раджабов С.С, Абдукадиров Б.А. / Шахсни биометрик идентификациялаш тизимларида сохта киришни аниқлаш муаммоси // Мухаммад ал-Хоразмий авлодлари, Ташкент 2020. — №3(13). - Б. 16–23. 2. Duc N.M. Your face is not your password / Black Hat Conference. 2009. – C. 1-16. 3. Chingovska, I. On the Effectiveness of Local Binary Patterns in Face Antispoofing / I. Chingovska, A. Anjos, S. Marcel // Biometrics Special Interest Group, 2012 BIOSIG - Proceedings of the International Conference of the.– 2012. 4. Maatta, J. Face spoofing detection from single images using micro-texture analysis / J. Maatta, A. Hadid, M.Pietikainen // Biometrics (IJCB), International Joint Conference on Biometrics, IEEE. – 2011. – С. 1-7. 5. Rekha, P.S. Spoofing Face Recognition Using Neural Network with 3D Mask / / IJETCSE. – 2015. – Т. 14 – № 1 – С. 123-127. 6. Yang, J. Face liveness detection with component dependent descriptor / J. Yang, Z. Lei, S. Liao // Biometrics (ICB). – 2013. – С. 1-6. 7. Способ идентификации личности по радужной оболочке глаза (варианты) : пат. 2407435 Рос. Федерация : МПК A 61 B 3/10 / Д. Е. Антонов ; заявитель и патентообладатель Антонов Дмитрий Евгеньевич. – № 2009128069/14 ; заявл. 22.07.2009 ; опубл. 27.12.2010. 8. Adamiak, K. Liveness detection in remote biometrics based on gaze direction estimation / K. Adamiak, D. Zurek, K. Slot // Proc. Fed. Conf. Comput. Sci. Inf. Syst. – 2015. – C. 225– 230. 9. Kant, C. Fake Face Recognition Using Fusion of Thermal Imaging and Skin Elasticity / C. Kant, N. Sharma // IJCSC. – 2013. – Т. 4 – C. 65-72. 10. Chen, S. Sensor-assisted facial recognition: an enhanced biometric authentication system for smartphones / S. Chen, A. Pande, P. Mohapatra // MobiSys '14. – 2014. – C. 109-122. 11. Ручай, А.Н. Модель атак и защиты биометрических систем распознавания диктора / А. Н. Ручай // Доклады ТУСУРа. – 2011. – № 1 – С. 96–100. 12. Ratha, N.K. Enhancing security and privacy in biometrics-based authentication systems / N. K. Ratha, J. H. Connell, R. M. Bolle // IBM Syst. J. – 2001. – Т. 40, № 3 – С. 614–634. 13. Yang, J. Learn convolutional neural network for face antispoofing. arXiv preprint arXiv:1408.5601, 2014 14. Li, L. An original face anti-spoofing approach using partial convolutional neural network. In IPTA, 2016. 15. Xu, Z. Learning temporal features using LSTM-CNN architecture for face anti-spoofing. ACPR, 2015, pp. 41-45 16. Rodriguez P., A painless attention mechanism for convolutional neural networks, ICLR, 2018, pp. 1-8