V-PVP readout boosts AI video detection to 95.28% AUC
A 0.5M parameter module unlocks frozen video backbones for fake video detection.
AI-generated videos often contain subtle temporal artifacts from inter-frame inconsistencies, yet standard video backbones with global readouts compress each frame into a single descriptor, suppressing the patch-level dynamics needed for detection. This explains why video pretrained backbones sometimes fail to outperform image-based probes. The authors identify excessive spatiotemporal aggregation as the core issue.
Their solution, Velocity Gated Patch Velocity Profiling (V-PVP), replaces the aggregation layer with two parallel streams over the patch velocity field, adding only about 0.5M trainable parameters. V-PVP is a plug-and-play module that consistently improves performance across diverse video backbones in both fine-tuning and linear probing settings. It reaches 95.28% AUC on AIGVDBench with the backbone frozen, demonstrating that simply changing the readout unlocks the temporal potential of pretrained video models for AI-generated video detection.
- Standard video backbones fail at AI video detection due to excessive global spatiotemporal aggregation that discards patch-level inter-frame dynamics.
- V-PVP adds only ~0.5M parameters and replaces the readout with two parallel streams over the patch velocity field, preserving local temporal cues.
- Achieves 95.28% AUC on AIGVDBench with frozen backbone, outperforming strong image-pretrained probes and restoring video backbones' advantage.
Why It Matters
Practical, low-cost upgrade to catch AI-generated videos without retraining entire models—critical for deepfake detection at scale.