International Journal of Science, Engineering and TechnologyAuthors: Er. Sunil Rathore, Assistant Professor Er. Laxmi Narayan Balai
Abstract: The rapid development of generative artificial intelligence has made manipulated videos increasingly realistic and difficult to verify through visual inspection alone. Deepfake videos may imitate a person’s identity, facial movements, expressions, or speech, creating risks for journalism, digital communication, privacy, and evidentiary media. This paper presents a feature-comparison framework for identifying suspicious temporal changes in video. The proposed workflow extracts frames from an input video, normalizes them, selects a reference frame, and compares subsequent frames using facial landmarks and object-level visual features. The principal indicators include facial landmark positions—especially around the eyes and mouth—colour or hue, contrast, and the relative arrangement of visible objects. Individual feature similarities are combined into an overall comparison score. A threshold such as 90 on a 0–100 scale may be used to flag frames for further inspection; it is a design parameter rather than a validated universal decision boundary. The framework is intended to support explainable, frame-level screening and localization of suspicious regions, not to replace expert verification. The source thesis identifies datasets such as FaceForensics++, DFDC, Celeb-DF, UADFV, and DeeperForensics-1.0 as potential evaluation resources. However, performance claims require reproducible testing on clearly defined train/test splits and diverse video conditions. The paper outlines the motivation, related approaches, proposed pipeline, evaluation plan, limitations, and future directions.
