V-FIND: Uncovering Sparse Forgery Neurons in Video Detectors

A new research paper on arXiv (ID: 2608.03008) proposes a framework called V-FIND (video forgery-intrinsic neuron discovery) to improve video forgery detection by identifying and activating sparse forensic knowledge within existing detectors, rather than retraining full models. The study finds that forgery-discriminative knowledge is concentrated in a sparse set of functionally specialized neurons, not uniformly distributed. V-FIND first localizes critical layers showing pronounced discrepancies between real and forged videos, then identifies the specific neurons responsible. This approach aims to reduce resource-intensive full-model retraining while maintaining or improving detection performance. The paper is categorized as a cross-announcement and is available at https://arxiv.org/abs/2608.03008.