Light-wDEF: lightweight deepfake evaluation framework for efficient and reliable visual detection

Abstract

Deepfake technology synthesizes and manipulates digital content to produce artificial videos in which an individual’s identity, facial appearance, or actions are convincingly altered. These forgeries can be highly realistic, posing serious risks to trust in digital media and undermining personal reputation, dignity, and the integrity of public discourse. This paper presents the Lightweight Deepfake Evaluation Framework (Light-wDEF)– a resource-efficient, reproducible framework tailored for large-scale evaluation of visual deepfake detectors on modest hardware. Light-wDEF combines deterministic frame sampling, batched MTCNN face extraction, streaming append-only storage with lazy offset indexing, online Albumentations augmentations, stratified splits with class-aware sampling, and robust training (AdamW, optional focal loss, per-epoch checkpointing with exact resume). Light-wDEF is designed as a reproducible and resource-conscious evaluation framework that focuses on experimental standardization, reproducibility, and efficient benchmarking of deepfake detection models. Using Light-wDEF, ConvNeXt-Base is identified as the most effective visual model in this study on the basis of AUC, balanced accuracy, F1, precision and recall while remaining compatible with Colab T4-class resource constraints. Ten modern visual models were trained and evaluated under identical conditions across five public datasets (Celeb-DF, FaceForensics++, Google DFD, FakeAVCeleb, DFDC Preview). Results show that convolutional families provide the best practical balance of discrimination and efficiency (ConvNeXt-Base achieves AUC≥0.96\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$AUC\ge 0.96$$\end{document} and F1≥0.94\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {F1}\ge 0.94$$\end{document} on the evaluated sets). The findings also expose limitations of visual-only detection on multimodal corpora such as FakeAVCeleb, motivating future multimodal fusion and domain-adaptation research. Fixed seeds, archived checkpoints, and environment metadata are provided to support exact replication.