Sparse Prototype-Based Framework for Deepfake Detection Using Class-Specific Dictionary Learning
Fatima Khalid, Usman Haider, Muhammad Iftikhar Hanif, Ahmar Rashid, Akhtar Khalil · IEEE Access · 2025
Deepfake detection remains a pressing challenge due to the rapid evolution of forgery techniques and the demand for robust, generalizable, and interpretable solutions. We present a sparse prototype-based framework for deepfake detection that integrates class-specific dictionary learning with high-capacity face embeddings from the InceptionResNetV2 backbone. Discriminative dictionaries for real and manipulated faces are learned, and Orthogonal Matching Pursuit is employed to encode sparse representations. Classification decisions are derived from reconstruction residuals and sparse coefficient patterns, enabling transparent, traceable predictions. The proposed method is evaluated on Celeb-DF and five FaceForensics++ manipulation subsets, including FaceShifter, NeuralTexture, DeepFakes, FaceSwap, and Face2Face. It achieves up to 98.21% accuracy on FaceSwap and 92.05% on Celeb-DF, demonstrating competitive performance with state-of-the-art deep architectures while maintaining moderate computational cost. Beyond accuracy, the framework offers intrinsic interpretability by linking predictions to semantically meaningful prototypes, making it particularly suitable for forensic and high-stakes verification scenarios.