Hybrid Multimodal Fake-News Detection Platform

Misinformation rarely gives itself away in the text alone. The wording, the image and the way a story spreads each carry part of the evidence, and a classifier that forces every case into two labels hides its own uncertainty.

• A hybrid classifier fusing text, image-forensics and propagation-pattern signals.
• Three output categories — Real, Fake and Unverified — instead of a forced binary.
• A FastAPI inference API returning a verdict, a calibrated confidence score and a class-probability breakdown.
• A threshold policy that routes low-confidence predictions into a manual review queue.
• An explainability panel surfacing the tokens and modalities driving each decision.
• A Streamlit analyst dashboard on top of the API.

• Built the fusion model and the inference service.
• Designed the confidence-threshold policy and the human-in-the-loop correction workflow.
• Captured reviewer verdicts as labelled data for retraining.

• A third “Unverified” label so the model can express uncertainty instead of guessing.
• A confidence threshold decides what the system labels itself and what a person reviews.
• Reviewer corrections feed back into the dataset as labelled examples.

• Each verdict ships with a confidence score, a probability breakdown and an explanation of what drove it.
• Low-confidence predictions are reviewed by a person rather than published as fact.