Advancements in Deepfake Detection: A Systematic Review and a Hybrid AI-Blockchain Framework Proposal

Abstract

Introduction:Despite the initial intent of deepfake technology for creative purposes, the need for deepfake detection continues to grow since technologies designed to mislead undermine digital trust through misinformation, identity theft, and an uninformed level of manipulation or misrepresentation of people in photographs, videos, social media and other contexts/overtures. The challenges raised by deepfakes - and deepfake detection - are intertwined with the United Nations Sustainable Development Goals (SDGs), especially SDG 16: Peace, Justice and Strong Institutions - integrity and accountability; SDG 9: Industry, Innovation and Infrastructure - secure, accessible, resilient and sustainable digital ecosystems; and SDG 4: Quality Education - digital literacy to counter disinformation, misinformation, deception, and manipulation.

Classic forensic implements to examine for facial inconsistencies or eye blinking observed in different temporal windows are already inadequate as deep learning–based generation methods improve practically on a daily basis. Work has examined CNN, RNN, Vision Transformers, GAN-artifact analysis, and other methods, yet no investigation produced a single approach with universal reliability. This review provided a survey of what deepfake detection currently looks like in terms of dominant methods employed in studies, their respective strengths and weaknesses, and discussed innovations in deepfake detection utilizing techniques such as blockchain or multimodal approaches. This review links the extensive use of deepfakes and rapidly contributing to advancing the broader sustainability agenda and calls for a sustainable systems-level change to ensure accurate, ethical, and trustworthy media authentication systems.

Objectives: The main aims of this study are to explore deepfake detection methods, critique their architectures, detection rates, weaknesses and strengths and their relative ability to detect deepfakes. In addition, the study proposes a holistic detection framework by exploring an ensemble of deep learning models with a blockchain-based verification system to enable trust, reliability and scalability. Lastly, the study also offers future research agenda that can support enhancing detection of deepfakes and media authentication.

Methods: To accomplish this, the proposed methodology proposes to use a multi-branch deep learning architecture. CNNs will be used for local feature and inconsistency detection at the pixel level while Swin Transformers will be used to capture global contextual patterns and dependencies. A GAN artifact-detecting component will also identify minor generative artifacts. Lastly, a blockchain layer for logging detection results will be included to protect the integrity of the results and provide tamper-proof confirmations. Altogether, this architecture leverages the varied strengths of the models used in novel ways while mitigating trust concerns with the help of a decentralized verification process.

Results: The use of this hybrid model has a number of important advantages. It has better detection reliability than single models and shows more resilience to emerging deepfake generation methods. The system can be deployed at scale and applied to real-world scenarios. Most importantly, blockchain technology combines the ability to have a permanent and confirmable record of detection results that improves the transparency and trust in the system.

Conclusions: The study concludes that no single detection method can effectively address the challenges posed by deepfake technology and hybrid approaches are essential for building robust systems. The proposed framework, which integrates CNNs, Swin Transformers, GAN artifact detectors and blockchain verification, presents a more powerful and trustworthy solution to the problem. Future directions for research include leveraging federated learning to preserve data privacy, implementing Zero-Knowledge Proofs for secure validation and advancing real-time multimodal detection techniques. Overall, this work lays a strong foundation for the development of next-generation deepfake detection and media authentication systems.