PRISM: a physics-reality integrated signal multistream framework for explainable deepfake detection using handcrafted forensic features

Deepfake detectors achieve strong benchmark performance but remain difficult to interpret and can degrade under unseen manipulation methods, compression and acquisition conditions. We present PRISM (Physics-Reality Integrated Signal Multistream) a video-level framework based on 50 explicit handcrafted forensic descriptors capturing sensor-residual, physiological, geometric, motion, texture, illumination and compression evidence. MediaPipe is used only as a fixed pretrained landmark-localization utility; the forensic representation contains no learned deep visual embeddings, and classification is performed from the resulting 50-dimensional video representation. Under an identity-disjoint FaceForensics + + c23 protocol, PRISM-50 achieved AUCs of 0.9706, 0.8096, 0.9631 and 0.7867 for DeepFakes, Face2Face, FaceSwap and NeuralTextures, respectively. Zero-shot Celeb-DF v2 evaluation yielded an AUC of 0.6322, with substantial shifts in compression, noise, physiological and geometric descriptors. Across five recent DF40 manipulation pipelines, zero-shot AUC ranged from 0.5995 to 0.9719, demonstrating strong generator dependence. Logistic Regression, Random Forest and LightGBM retained useful discrimination from the same representation, whereas Xception and LSDA achieved higher predictive performance. SHAP stability and perturbation analyses showed that highly attributed descriptors materially influenced classifier decisions without implying causal evidence of manipulation. PRISM is therefore positioned as an interpretable forensic framework for examining measurable evidence, feature interactions and failure mechanisms across manipulation and acquisition domains.