Abstract
DeepFake-O-Meter v2.0 is an open-source, multi-modal platform designed to democratize access to state-of-the-art DeepFake detection. By integrating 18 diverse algorithms for image, video, and audio analysis into a unified web interface, the platform serves as a "one-stop shop" for journalists, researchers, and the general public. This paper details the system's architecture, its fair-scheduling mechanisms, and usage trends observed during its initial deployment.
Introduction
The proliferation of AI-generated content, or DeepFakes, poses a growing threat to digital trust, impacting media, politics, and security. While the research community has produced numerous effective detection models, these tools are frequently inaccessible to the public due to complex installation requirements, software dependencies, or commercial barriers. DeepFake-O-Meter v2.0 addresses this by providing a centralized, non-profit hub where users can analyze media using multiple top-tier detectors simultaneously.
System Architecture and Methods
The platform's architecture is split into a Python Flask/Apache front-end and a high-performance back-end. To overcome "dependency hell"—the conflict between different software libraries required by various AI models—the system utilizes Docker containerization. Each of the 18 integrated detectors operates within its own isolated environment.
To manage the computational load on its eight A5000 GPUs, the platform employs a priority-based scheduler. This "fairness algorithm" prioritizes low-frequency users, ensuring that a single power user cannot monopolize the system's resources. Currently, the platform hosts 6 image detectors (e.g., CLIP-ViT, NPR), 7 video detectors (e.g., FTCN, SBI), and 5 audio detectors (e.g., RawNet3, Whisper).
Integrated Detectors and Data
The platform hosts a wide array of detectors trained on diverse datasets, including StyleGAN2, DALL-E, and various face-swap and audio spoofing sets. Beyond detection, the platform includes an authorized data collection module, allowing users to contribute samples to help researchers develop more robust future models.
Platform Performance and Insights
Analysis of 115 days of usage (Feb–June 2024) revealed significant demand, with over 600 registered users from 98 countries submitting more than 4,000 tasks. Image and audio tasks typically complete within 30 seconds, while video analysis takes approximately 90 seconds due to the complexity of frame-by-frame processing.
A key feature of the platform is the "Detailed Result Page," which provides not only the detection probability but also references to the underlying academic papers and source code. This transparency transforms the platform into a living literature review for practitioners, making academic research actionable for non-experts.
Conclusion and Future Work
DeepFake-O-Meter v2.0 successfully provides a standardized environment for benchmarking and public use. Future developments will focus on multi-modal detectors—identifying inconsistencies between audio and visual tracks—and detecting "out-of-context" images where AI-generated content conflicts with accompanying text.
Summary updated