University Honors

Deepfakes have become increasingly realistic with the rise of diffusion models, vision transformers, and multimodal generators, posing major risks to security, trust, and authenticity in digital media. While existing detectors achieve strong benchmark accuracy, their performance deteriorates under real-world distortions such as compression, transmission loss, or bitrate fluctuations. This literature review traces the evolution of deepfake detectionófrom early spatial and temporal CNNs to frequency-, latent-, and transformer-based architecturesóand examines robustness-oriented frameworks including ADD, QAD, BZNet, and DPL. To address the gap between laboratory conditions and deployment environments, this study introduces two complementary research directions. First, an adaptive detection framework, the MDN-Guided Noise-Aware Modular Ensemble, models test-time uncertainty using a Mixture Density Network to dynamically select noise-specific fine-tuning modules. Second, a new dataset, DeFND (Deepfake Forensics under Network Degradation), captures authentic Wi-Fi and cellular transmission artifacts by streaming deepfake videos through real WebRTC pipelines. Together, these contributions provide a foundation for evaluating and improving deepfake detectors under realistic network and corruption conditions, advancing the pursuit of reliable, uncertainty-aware detection systems for practical deployment.