The Impact of Deepfakes on Digital Media Authenticity
Authors: Aditi Indre, Jayesh Shinde, Dr. Srivaramangai Ramanujam
Published: IJRASET, 23 March 2025 · ISSN 2321-9653
DOI: 10.22214/ijraset.2025.67796
A survey of the deepfake detection landscape, focused on multimodal approaches that combine visual, audio, and physiological signals rather than relying on any single modality.
The paper reviews state-of-the-art detection models (CNNs, Vision Transformers, XGBoost classifiers) and their evaluation against the Celeb-DF and FaceForensics++ benchmarks. It covers facial landmark analysis, voice consistency checking, and heart-rate variability extraction as complementary detection signals.
The central argument is that generalization, not raw accuracy, is the open problem. Models perform well against the manipulation techniques they were trained on and degrade against novel ones, which makes adversarial training and adaptive approaches more important than incremental accuracy gains. Real-time processing and interpretability remain unsolved for forensic and legal use, where an unexplainable classifier isn’t useful regardless of its score.