The Regularizing Power of Language-Training Deepfake Detectors
Researchers have developed a novel method for training deepfake detectors by leveraging multimodal large language models (MLLMs). This approach uses language as a regularization mechanism to improve both the generalizability and interpretability of the detectors. The system employs a dual-encoder architecture and a two-stage training process, including reinforcement learning to encourage descriptive reasoning before classification, which significantly boosts performance and provides interpretable outputs. AI
IMPACT Enhances deepfake detection capabilities by improving generalization and interpretability, crucial for combating AI-generated misinformation.