Why Do AI Deepfake Detectors Fail on New Videos? | Research Gap Explained | Paper to Project #02

Can an AI deepfake detector score highly on one dataset—and still struggle with completely new videos? In Paper to Project #02, ADR Lab explores a 2026 research paper on deepfake detection and turns its findings into a practical, testable AI/ML project idea. This episode is intended for BTech, MTech, PhD scholars, AI/ML students, and researchers who want to learn how to read research papers, understand experiments, identify potential research opportunities, and plan reproducible projects. 📄 RESEARCH PAPER DISCUSSED Title: Deepfake Detection Across Image, Video, and Audio: A Comprehensive Survey With Empirical Evaluation of Generalization and Robustness Authors: Nguyen-Le et al. Journal: Artificial Intelligence Review (Springer Nature), 2026 DOI / paper: https://doi.org/10.1007/s10462-026-11608-4 🔍 IN THIS VIDEO • Why high within-dataset accuracy may not translate to unseen videos • How within-domain, cross-dataset, and robustness evaluation differ • What the authors found about generalization and adversarial robustness • Why compression and changes in data sources matter when testing detectors • How to turn the paper into a focused, realistic student research question • How to select datasets, baselines, experiments, and evaluation metrics • Why a potential research gap must be verified before claiming novelty 💡 THE PROPOSED RESEARCH DIRECTION Robustness-Aware Cross-Dataset Deepfake Detection ADR Lab proposes an evaluation-focused project: compare baseline deepfake detectors on held-out and separate datasets, test selected video transformations such as compression or resolution changes, and report video-level performance transparently. Candidate datasets include FaceForensics++ and Celeb-DF, subject to their access conditions and licenses. Possible evaluation measures include ROC-AUC, accuracy, false-positive and false-negative rates, and uncertainty where practical. The aim is not to claim a new detection algorithm but to design a clear, reproducible comparison that can be assessed against recent research. ⚠️ RESEARCH INTEGRITY NOTE This is a project proposal inspired by the published paper—not a completed ADR Lab experiment, and not a claim that the research direction is already novel. The video distinguishes the paper's results from our suggested student experiments. Verify dataset permissions and survey the latest related work before finalizing a novelty claim. 🎓 WHO SHOULD WATCH? BTech / BE final-year students • MTech / ME students • PhD scholars • Deepfake detection researchers • Computer vision and cybersecurity learners • AI/ML project developers 📚 SOURCE Research paper: https://doi.org/10.1007/s10462-026-11608-4 ▶️ PAPER TO PROJECT BY ADR LAB Our approach: Research Paper → Problem → Methodology → Evidence → Potential Gap → Proposed Experiment → Novelty Check 💬 Which part of deepfake detection should we explore next: cross-dataset testing, video compression, or explainability? Comment below. 👍 Like this video if it helps your research journey, and subscribe to ADR Lab for more research-paper-to-project discussions. 🔜 COMING NEXT — EPISODE #03: Explainable AI in Healthcare #PaperToProject #DeepfakeDetection #AIResearch #MachineLearning #ResearchGap #ComputerVision #BTechProjects #MTechProjects #ADRLab #DeepLearning