Enterprise AI Analysis

Progressive Gated Co-Teaching for Weakly Supervised Deepfake Detection

Rui Lang, Guangsheng Yu, Qin Wang, Xu Wang — April 12-18, 2026

The surge of diffusion- and GAN-based video generators has produced photorealistic forgeries that are increasingly difficult to distinguish from authentic content under weak supervision. The Co-Teaching framework, which consists of two collaboratively trained networks that exchange pseudo-labels to mitigate label noise, has shown promise in localizing forged regions. However, it still suffers from early-stage noise amplification and unstable reciprocal supervision, especially when trained with only video-level labels. In this paper, we propose a Vision Transformer (ViT)-based dual-branch framework that progressively enhances weakly supervised deepfake localization. From both spatial appearance and temporal dynamics perspectives, the two ViT branches perform score-guided token condensation: a learned scorer ranks patch tokens and condenses them before any supervision, ensuring gradients focus on discriminative evidence rather than diffuse background. To stabilize co-learning under noisy labels, we introduce a progressive co-teaching mechanism that integrates Exponential Moving Average (EMA) smoothing and gated token exchange. The EMA teachers provide temporally smoothed predictions that suppress transient fluctuations, while the gated token exchange, which includes confidence and consensus gates, selectively filters unreliable cross-branch supervision. Together, these mechanisms make supervision explicit in both timing ("when") and scope ("what"), yielding smoother and more reliable optimization. Experiments demonstrate that our framework achieves more stable convergence and accuracy than existing co-teaching and transformer baselines. Ablation studies further confirm that token selection before supervision and progressive, gated exchanges are key to improving both robustness and generalization.