Privacy-focused artificial intelligence model for detecting deepfake-based cyber threats

This paper presents an extended literature survey on deepfake-oriented cyber threats with a strong focus on practical deployment constraints. Although detection accuracy has improved in recent years, real-world adoption remains limited by privacy concerns, hardware requirements, and weak generalization across unseen media conditions. We examine the evolution of deepfake detection from handcrafted forensic cues to deep multimodal architectures, and we discuss why many high-scoring benchmark models fail under operational workloads. The survey highlights a local-first strategy that combines multimodal evidence, interpretable outputs, and resource- aware model design so that robust detection can run on consumer-grade systems. The goal is to support trustworthy, privacy-preserving defense against impersonation fraud, misinformation, and identity abuse in modern digital ecosystems.