AI-Based Deepfake Scam Detection for Protection Against Voice and Video Fraud

The proliferation of deepfake technology, under-pinned by Generative Adversarial Networks (GANs) and diffusion-based synthesis pipelines, has created an escalating cybersecurity crisis. These systems produce synthetic audio-visual content of sufficient fidelity to deceive human observers, and are increasingly leveraged in financial fraud, identity imper-sonation, and coordinated disinformation. This work delivers a structured technical survey of AI-driven detection methodologies spanning visual, acoustic, and behavioural modalities. Meth-ods are systematically compared with respect to architecture, benchmark performance, and operational limitations. Drawing on insights extracted from this survey, we formulate a conceptual three-stream detection architecture that unifies spatial artifact analysis, spectral audio verification, and temporal behavioural modelling via a Transformer-based cross-modal attention fusion mechanism. Our analysis demonstrates that multi-modal fusion consistently surpasses single-stream detection and exposes crit-ical open problems, including generalisation under distribution shift, robustness to adversarial manipulation, and edge-device deployment constraints.