Technologie rozpoznání obličeje vs. Deepfake: analýza bezpečnostních rizik a protiopatření

Abstract

This thesis analyzes the security risks associated with face recognition (FR) technologies and the synthesis of facial identities using so-called deepfake techniques. The aim of the work is to design and implement a modular system that combines identity recognition with manipulation detection in video content. The theoretical part describes the principles of FR systems, provides an overview of generative tools for creating deepfake content, and presents the latest detection methods, including convolutional and transformer-based models. The practical part focuses on evaluating selected FR models (FaceNet, VGG-Face, InsightFace) in terms of accuracy and robustness when confronted with forged content. For deepfake manipulation detection, modern architectures such as XceptionNet, ResNet152V2, EfficientNetB3, and Vision Transformer were tested. The results were evaluated using metrics such as FAR, FRR, and EER. The thesis also includes the development of a web application enabling real-time detection of deepfake manipulations, video processing, and identity database management. The conclusions highlight the limitations of current detection methods in the context of increasingly realistic deepfake content and emphasize the need to combine multiple approaches to enhance the resilience of biometric systems.