Reproduce “Tutor-Student Reinforcement Learning: A Dynamic Curriculum for Robust Deepfake Detection” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhanhe Lei, Zhongyuan Wang, Jikang Cheng, Baojin Huang, Yuhong Yang, Zhen Han, Chao Liang, Dengpan Ye

The Tutor-Student Reinforcement Learning (TSRL) framework introduces an adaptive curriculum where a Reinforcement Learning agent dynamically weights training samples for deepfake detectors. This approach consistently enhanced generalization across various architectures, improving average AUC by up to 4.1% on unseen manipulation techniques and data domains.

Key insights

  • An RL-driven dynamic curriculum significantly outperforms static, heuristic-based curriculum learning methods, leading to substantially improved generalization in deepfake detection.
  • The adaptive sample weighting strategy enables the Student detector to learn more discriminative and disentangled feature representations, clearly separating real from fake data and organizing fake samples by their learning difficulty.
  • The framework effectively guides the student model to prioritize and focus its learning capacity on challenging examples near the decision boundary by isolating trivial samples.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.