Reproduce “Robust Deepfake Detection: Mitigating Spatial Attention Drift via Calibrated Complementary Ensembles” on OpenResearch

About this paper

Minh-Khoa Le-Phan, Minh-Hoang Le, Trong-Le Do, Minh-Triet Tran

A deepfake detection framework enhances robustness against real-world media degradations by employing an extreme compound degradation engine and a multi-stream architecture. This approach achieved an AUC of 0.8523 on the private test set in the NTIRE 2026 Robust Deepfake Detection Challenge, effectively mitigating spatial attention drift and texture bias in detection models.

Key insights

• Extensive data diversity, combined with aggressive compound degradation during training, is crucial for achieving stable zero-shot generalization and maintaining robustness against real-world noisy media.
• Parameter-efficient tuning (LoRA) of large foundation models (DINOv2) is more effective than full fine-tuning for preserving generalized priors and cross-domain robustness in deepfake detection.
• A complementary multi-stream architecture, with each stream specialized for different types of features, effectively mitigates spatial attention drift and texture bias, confirming that individual streams learn distinct and non-redundant features.

Summary by alphaXiv.