Practical Deepfake Detection: Vulnerabilities in Global Contexts

Detection systems fail on low-quality videos: State-of-the-art deepfake detection algorithms remain vulnerable to video quality degradation, meaning they may incorrectly flag authentic videos as fake when they're compressed or corrupted—a critical problem in global contexts where internet bandwidth limitations are common.

Training-test mismatch undermines real-world performance: While detection models perform well against corruptions similar to those used during training, they struggle with real-world video degradation, revealing a significant gap between laboratory conditions and practical deployment that threatens the reliability of these systems.

Real-world consequences are already happening: The paper cites the case of Gabonese President Bongo's address, where a detection algorithm authenticated the original video but flagged degraded versions as fake—demonstrating that current vulnerabilities can have immediate political and social consequences.

Global inequality amplifies the problem: The paper emphasizes that practical deepfake detection must account for "global contexts," suggesting that regions with lower video quality infrastructure may face disproportionate risks of authentication failures, potentially undermining trust in legitimate content.

Technical robustness alone is insufficient: The authors highlight both "technical and ethical avenues of exploration," indicating that solving deepfake detection requires more than just improving algorithms—it demands consideration of how these systems will actually be used and who they might harm in practice.