TruthLens: Training-Free Data Verification for Deepfake Images via VQA-style Probing

Ritabrata Chakraborty, Rajatsubhra Chakraborty, Ali Khaleghi Rahimian

Keywords: Data Verification, Deepfake Detection, Instance-Level Auditing, Generative Models, Multimodal Probing

TL;DR: TruthLens is a training-free, multimodal-probe pipeline that performs instance-level data verification on images produced by large generative models, flagging deepfakes and explaining its verdicts to advance data-centric safety.

Abstract: AI-generated imagery is on the rise to be outplacing our human ability to spot manipulations. Prevailing deepfake detectors cast as opaque binary classifiers offer little to no insights into their decisions. We introduce TruthLens, a training-free framework that reframes deepfake detection as a VQA task. We leverage large vision-language models (LVLMs) to reveal artifacts and GPT-4 to reason over the evidence to reach a coherent verdict by fusing visual and semantic cues. The framework explains which artifacts triggered its judgement, providing a deeper and newer mode of transparency. Evaluations demonstrate that TruthLens outperforms conventional methods while maintaining a strong emphasis on explainability and delivering instance-level data verification for large-scale generative models.

Submission Number: 3