Deepfake Detection using TCN and EfficientNet-B3
Year: 2025
Type: article
Abstract: The increasing computational capabilities ofmodern systems have significantly enhanced deep learningalgorithms, making it easier to create highly realistic, AI-generated videos, commonly known as deepfakes. Thesesynthetic videos, which seamlessly replicate humanappearances, pose serious threats, including politicalmanipulation, fabricated terrorist events, revengepornography, and blackmail. In this study, we propose anovel deep learning-based approach for effectivelydistinguishing between authentic and AI-generated fakevideos. Our method is designed to identify both facereplacement and reenactment deepfakes. Leveraging thepower of Artificial Intelligence (AI) to counter the misuse ofAI, our system utilizes an EfficientNet convolutional neuralnetwork to extract frame-level features. These features aresubsequently used to train a Temporal ConvolutionalNetwork (TCN) to classify videos as either real ormanipulated. To ensure robust evaluation and emulate real-world scenarios, we employ the well-known DeepfakeDetection Challenge[1] dataset for training and testing. Ourproposed system aims to achieve competitive results througha simple yet effective approach.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4409696115