This paper addresses the problem of deepfake detectors trained in controlled environments failing to generalize when applied to real-world environments due to the proliferation of AI-generated videos on social networks. Aggressive proprietary compression on platforms like YouTube and Facebook is identified as a major cause of the removal of low-level forensic clues. Given the difficulty of replicating these transformations at scale due to API restrictions and data sharing constraints, this paper proposes a framework that emulates the video sharing pipeline of social networks by estimating compression and resizing parameters from a small number of uploaded videos. Using these parameters, we implement a local emulator capable of reproducing platform-specific artifacts on large datasets without direct API access. Experimental results demonstrate that the emulated data closely resembles the degradation patterns of real-world uploaded videos, and that detectors fine-tuned on emulated videos achieve comparable performance to detectors trained on real-world shared media. This research provides a scalable and practical solution that bridges the gap between lab-based training and real-world deployment, particularly in the unexplored area of compressed video content.
