EVALUATION OF METHODS FOR DEEPFAKE DETECTION IN VIDEO CONTENT

Evplov Nikita Aleksandrovich

The growing sophistication and accessibility of deepfake technologies pose a serious threat to information security by facilitating the spread of disinformation and undermining trust in digital content. Effective countermeasures require the development and deployment of robust detection systems capable of operating under large data volumes and a variety of falsification techniques. This study presents a comparative evaluation of three distinct approaches to deepfake detection in video content—spectral analysis, a convolutional neural network (CNN), and a hybrid algorithm—with the goal of identifying the optimal balance among precision, recall, and performance. For the experiment, a representative dataset of over 4,000 video clips was compiled, including genuine recordings captured under various conditions and deepfakes generated using popular tools such as DeepFaceLab. The effectiveness of each method was assessed using standard metrics—precision, recall, and F1-score—as well as false positive and false negative rates. Additionally, algorithm performance was measured across different data volumes to evaluate scalability. Results demonstrated that the hybrid algorithm achieved the highest performance, with 95.8% precision and a 94.7% F1-score, indicating its superior ability to detect forgeries while minimizing errors. The CNN offered a balanced solution, only slightly trailing in precision but surpassing the hybrid method in processing speed. Spectral analysis proved to be the fastest yet least accurate approach. It was also observed that as deepfakes became more realistic and system load increased (through parallel stream processing), the effectiveness of all methods declined; however, the hybrid approach remained the most robust. The study confirms that no universal solution exists for deepfake detection. The choice of the optimal method depends on specific requirements: the hybrid algorithm is preferable when maximum accuracy is needed, whereas spectral analysis and CNNs are better suited for real-time applications. The most reliable defense strategy involves multi-layered systems combining several methods, thereby enhancing overall resilience against continually evolving video falsification technologies.

Keywords:deepfake, detection methods, convolutional neural network, video content, performance evaluation