Browse Teach Curate

Interactive Deepfake Analysis with Instruction-Tuned Multimodal Language Models

Recasts deepfake analysis from a discriminative classifier into a conversational multimodal-LLM task that detects forgery, classifies its technique, describes visual artifacts, and answers follow-up questions. Introduces a GPT-assisted instruction-following dataset, an evaluation benchmark, and a LoRA-fine-tuned baseline system for interactive forgery investigation.

Dr. Theopolis Text

Towards Interactive Deepfake Analysis Lixiong Qin∗†, Ning Jiang∗, Yang Zhang∗, Yuhan Qiu∗, Dingheng

Most deepfake analysis systems are discriminative classifiers that only output a real/fake label or a forgery-type category, which limits how a human investigator can work with them. This paper refra…