From pixels to proof: explainable deepfake detection framework
Anita Khadka, Carsten R. Maple · IET conference proceedings. · 2025
Deepfake technologies are rapidly evolving, posing significant challenges to digital media authenticity as generative models become increasingly capable of producing highly realistic data (e.g, images, audio). From reputational damage to misinformation, the risks associated with undetected manipulated content are escalating. This paper presents an explainable deepfake detection framework that enhances detection accuracy and transparency, aiming to support a trustworthy digital infrastructure. We propose a novel approach fusing Vision Transformers (ViT) and High-Resolution Networks (HRNet) to extract both global semantics and fine-grained visual cues and combining them using multi-head self-attention and a meta-learning model that improves generalisation to diverse and evolving manipulation techniques. A key contribution of this work is the integration of explainability mechanisms to provide transparency and interpretability in model decision-making. We evaluate our method on benchmark image-based deepfake datasets (Flickr-Faces-HQ (FFHQ) and Synthetic Faces-HQ (SFHQ)) and demonstrate higher performance over selected existing models, such as ViT-only and HRNet-only, in terms of performance. This work contributes to the responsible use of Artificial Intelligence (AI) in media authentication, paving the way for future multimodal extensions that incorporate audio-visual and temporal cues.