INSIDER: Mining Latent Forensic Knowledge for AI-Generated Video Detection

As AI-generated videos become increasingly realistic, reliably distinguishing authentic from AI-generated content is essential for maintaining trust in visual information. However, existing learning-based detectors often rely on large amounts of labeled data for task-specific training, while latent forensic knowledge within pretrained video models remains underutilized. To address this limitation, we propose INSIDER, an AI-generated-video detection framework that uses a small set of labeled source videos to mine forensic knowledge within a frozen video encoder. Specifically, we first use layer selection to identify candidate layers based on the source-validation performance of layer-wise linear probes. Next, we combine decision leverage (DL) and supervised sensitivity (SS) to assess how pooled neuron responses influence fixed probe scores and how sensitive the same probe's supervised authenticity loss is to those responses for each source class. Forensic neurons are then selected adaptively using the knee points of layer-wise score curves. Finally, sparse readout concatenates the selected responses across layers into a compact forensic representation, on which we train a lightweight linear classifier for authenticity detection. With the detector fixed during target evaluation, INSIDER outperforms the evaluated detection methods on all four benchmarks covering 57 generator subsets, achieving a benchmark-averaged macro AUC of 92.53%. Further experiments show that INSIDER outperforms direct final-layer readouts with each of four video encoders. In separate tests, the source-trained detector remains robust to video compression and substantial resolution changes.