Naučni rad: "Adaptive method of biometric authentication, AC-CNN based on convolutional neural networks under conditions of partially available biometric data"

Abstract—Digital transformation and the rise of electronic payments in the post-COVID19 period require more accurate methods of biometric authentication. Standard biometric authentication methods show limitations when users wear masks and gloves. In this paper, the method of biometric authentication AC-CNN (Adaptive Clustered Convolutional Neural Network) is proposed. AC-CNN is based on the classification of facial images into clusters and their vectorization using previously trained CNN facial recognition models (e.g.VGGFace, ArcFace, DeepFace). After classification, the facial image is vectorized by the CNN algorithm, which shows the best facial recognition accuracy for facial images from that cluster. The registration and verification processes use different types of CNN algorithms. In the process of user identification, after successful user identification, the type of CNN algorithm that was used for that user in the registration process is loaded from the database and the same CNN algorithm will be used in the verification process. In the verification process, the biometric template of the identified user
is compared with the reference biometric template contained in the database and based on their Euclidean distance, a verification decision is calculated. Since the biometric templates of the users in the database are divided into clusters, an allowable decision threshold is calculated separately for each cluster. The results of experimental research show that the presented method allows for increased verification accuracy
even in conditions where the face is masked and when only partial biometric data is available.

Keywords—Biometric authentication, AC-CNN, Face recognition, masked face, partial biometric data

Introduction: In this paper, the method of biometric authentication AC CNN (Adaptive Clustered – Convolutional Neural Network) is presented. The AC-CNN model performs adaptive clustering of facial image features to improve verification accuracy. Each cluster labels the same class of facial images, for example, a white male face, or a black female face, and uses a type of pre-trained CNN algorithm, which has demonstrated high facial recognition accuracy performance for that face class. With this CNN algorithm, vectorization of images of faces classified in that cluster is performed. The method uses some of the most accurate pre-trained CNN algorithms available, such as VGGFace [1], OpenFace [2], DeepFace [3], ArcFace [4]. The use of this method is intended for extreme cases, such as when users wear a face mask and when complete biometric data is not available. The results of experimental research show improved performance compared to the reference E3FRM method [5]...

Published in: 2025 6th International Workshop on Engineering Technologies and Computer Science (EnT)
Date of Conference: 29. October 2025
Date Added to IEEE Xplore: 19 November 2025

ISBN Information: Electronic ISBN:979-8-3315-7253-2 , Print on Demand(PoD) ISBN:979-8-3315-7254-9
ISSN Information: Electronic ISSN: 2767-1976 , Print on Demand(PoD) ISSN: 2767-1984

DOI: 10.1109/EnT68818.2025.11245710
Publisher: IEEE
Conference Location: Sankt Peterburg, Russian Federation

Referenca:

N. Badovinac and D. Simic, "Adaptive method of biometric authentication, AC-CNN based on convolutional neural networks under conditions of partially available biometric data," 2025 6th International Workshop on Engineering Technologies and Computer Science (EnT), Sankt Peterburg, Russian Federation, 2025, pp. 1-7, doi: 10.1109/EnT68818.2025.11245710