Title:
Interpretable Texture, Frequency and Structural Features for Explainable Fake Image and Deepfake Detection
Authors:
Shailendra Singh Kathait (Co-Founder and Chief Data Scientist Valiance Solutions Noida, India)Ashish Kumar (Principal Data Scientist Valiance Solutions Noida, India)Dheeraj Kumar Prajapati (Intern Data Scientist Valiance Solutions Noida, India)
Riya Singh (Intern Data Scientist Valiance Solutions Noida, India)
Summary:
Generative models have significantly increased the realism of synthetic images, which makes reliable fake image detection a challenging problem. Deep learning detectors achieve high accuracy, but they are difficult to interpret and often generalize poorly to unseen generators. This paper presents a set of handcrafted, interpretable features that capture intrinsic differences between real and synthetic images from three complementary perspectives: texture complexity, frequency-domain repetition, and geometric consistency. Texture richness is quantified by Local Binary Pattern (LBP) entropy and local density, repetitive upsampling artifacts are measured through Fourier spectrum analysis, and structural consistency is measured through edge-based line statistics and structure-tensor coherence. Each feature is defined formally and analyzed in terms of computational cost, expected behavior, and known confounding factors, and an evaluation protocol covering image, video, cross-generator, and post-processing scenarios is specified. The features are deployed in a detection platform, where they provide feature-level explanations for the prediction of an image classifier, and a video pipeline aggregates frame-level decisions into a video-level verdict. The descriptors are computationally efficient, model-agnostic, and suitable for classical machine learning as well as hybrid detection frameworks.