A Comprehensive Survey of Attention-Based Deep Learning Techniques for Robust Deepfake Face Detection
Abstract
The rapid advancement of deep learning and generative artificial intelligence has significantly increased the realism of artificially generated and manipulated facial images, creating substantial challenges for digital media authenticity and security. Fake-face detection has emerged as an important research area for identifying synthetic faces, face-swapping operations, and manipulated facial content. Traditional detection approaches often struggle to identify subtle manipulation artifacts, particularly under variations in image quality, compression, lighting conditions, and facial expressions. This paper presents a comprehensive survey of attention-based deep learning techniques for robust fake-face detection, focusing on recent advancements in convolutional neural networks (CNNs), vision transformers, and hybrid deep learning architectures. The survey examines the role of spatial attention, channel attention, self-attention, and multi-head attention mechanisms in enhancing discriminative feature extraction and identifying subtle inconsistencies in facial regions. Furthermore, it reviews attention-based feature fusion strategies, transfer learning approaches, and multi-scale feature extraction techniques used to improve detection performance. Various benchmark datasets, evaluation metrics, and experimental methodologies are discussed to provide a structured understanding of existing research. The paper also investigates key challenges, including cross-dataset generalization, adversarial manipulation, computational complexity, interpretability, and the detection of highly realistic AI-generated faces. Finally, potential research directions are identified, including complementary attention mechanisms, frequency-domain analysis, lightweight detection architectures, explainable artificial intelligence, and multimodal learning. This survey aims to provide researchers with a comprehensive understanding of existing attention-based detection methods and support the development of accurate, efficient, and generalizable fake-face detection frameworks.