TLDR: A new framework helps deepfake detectors work better on social media videos by accurately mimicking how platforms like YouTube and Facebook compress and resize videos. This allows researchers to train detectors on realistic, compressed videos without needing direct access to social media APIs, overcoming a major hurdle in real-world deepfake detection. The study found that using at least 30 videos per resolution for emulation provides reliable results.
The rapid rise of AI-generated videos, commonly known as deepfakes, on social media platforms presents a significant challenge for detection technologies. While deepfake detectors perform well in controlled laboratory environments, they often struggle when faced with real-world videos shared on platforms like YouTube and Facebook. This performance gap is primarily due to the aggressive, proprietary compression and resizing techniques applied by these platforms, which strip away the subtle forensic cues that detectors rely on.
Replicating these complex compression transformations at scale has been difficult for researchers due to limitations in social media APIs and data-sharing restrictions. To address this critical issue, a new framework has been proposed that effectively emulates the video sharing pipelines of social networks.
Understanding the Challenge
Traditional deepfake detection methods, often based on convolutional neural networks (CNNs) or Vision Transformers (ViTs), analyze low-level features and inconsistencies introduced during the generative process. These methods achieve high accuracy under ideal lab conditions. However, when videos are uploaded to social media, platforms compress and resize them to save bandwidth and storage. This process degrades the video quality and, crucially, removes or alters the very forensic features that deepfake detectors are designed to identify. This means a detector trained on a pristine deepfake might fail to recognize it once it has been compressed and shared online.
Previous research has largely focused on detecting fake images, where API access for data collection is somewhat easier. The video domain, however, has remained significantly underexplored due to larger file sizes, higher processing demands, and more stringent API limitations.
The Emulation Framework: Bridging the Gap
The proposed Social Network Video Sharing Emulator (SNVSE) offers a novel solution. It works by first estimating the specific compression and resizing parameters used by a target social network (e.g., YouTube or Facebook). This is achieved by uploading a small set of fewer than 50 videos per resolution to the platform. By comparing the original videos with their shared versions, the framework can infer crucial re-encoding parameters, such as the Compression Rate Factor (CRF) and target resolutions. These parameters are then stored in a local database.
Once these parameters are established, the framework can then apply them to any large dataset of non-shared videos. This means researchers can take their existing deepfake video datasets and process them locally to accurately mimic the degradation patterns of real social media uploads. This local emulation bypasses the need for direct API access, avoiding associated costs, rate limits, and content-sharing quotas.
Experimental Validation and Key Findings
The researchers tested their framework using the FaceForensics++ (FF++) dataset, which includes both original and social media-shared videos. They trained several popular deepfake detection models (DenseNet, InceptionNet, XceptionNet, and ResNet-50) under different conditions:
- Trained on non-shared videos.
- Fine-tuned on actual social media-shared videos.
- Fine-tuned on videos processed by the new emulation framework.
The results were compelling. Models trained solely on non-shared videos showed a significant drop in their ability to detect fake videos once those videos were shared on social media. However, when these models were fine-tuned on either actual shared videos or the emulated videos, their performance recovered substantially and comparably. This demonstrates that the emulated data is realistic enough to effectively improve detector performance in real-world scenarios.
The study also extended its evaluation to videos shared on Facebook, YouTube, and BlueSky in 2025, confirming the framework’s adaptability to current platform encoding pipelines. An important practical insight emerged from an ablation study: to ensure stable and reliable CRF estimation, at least 30 shared videos per resolution are required for the initial parameter estimation phase.
Also Read:
- DPGNet: Advancing Deepfake Detection with Unlabeled Data
- Enhancing Video Behavior Recognition with a Hybrid AI Approach
Looking Ahead
This framework represents a significant step forward in making deepfake detection more robust and practical for real-world deployment. By providing a scalable and cost-effective way to generate realistically degraded video data, it helps bridge the gap between laboratory research and the challenges posed by social media compression. The researchers plan to expand the framework to estimate additional encoding parameters like frame rate and pixel format, and to include a wider range of platforms such as TikTok, X (formerly Twitter), Telegram, WhatsApp, and Slack.
For more technical details, you can refer to the full research paper: Bridging the Gap: A Framework for Real-World Video Deepfake Detection via Social Network Compression Emulation.
