Verify Any News Article
Upload text and an image to detect AI-generated content, manipulated images, and out-of-context pairings.
- Real news: paste a recent news article + its original photo.
- OOC: a real photo + a caption from a different unrelated story.
- Manipulated: a deepfake / face-swap image + plausible caption.
- AI-Text: a real photo + LLM-generated misinformation text.
- Fully fabricated: an AI-generated (MidJourney) image + AI-generated text.
Binary AI-image detector -- Approach 1 (fine-tuned RGB ResNet50 with Fourier masking). Retrained on all 8 GenImage generators with official ImageNet nature (overall test AP ~0.99). Upload an image to get P(AI-generated).
Binary AI-image detector -- Approach 2 (Dual-DCT ResNet50, frequency-domain input). Retrained on all 8 GenImage generators with official ImageNet nature. Upload an image to get P(AI-generated).
About
5-class multimodal fake-news detector (V3.1 spec, honest-path retrain) + two standalone binary AI-image forensic detectors (8/8 GenImage generators, official ImageNet nature). Stack: FND-CLIP semantic + DCT-Forensic image + Qwen2-7B text -> V3PairwiseFusion -> MLP.
Headline numbers: 3-seed ensemble test F1 = 0.7149, MMFakeBench transfer F1 = 0.7197 (+ bias correction). Full repo: https://github.com/FerasMad/MultiGuard