BertAndDeberta + ViT — Transformer Ensemble for Fake News Detection

This repository hosts a multimodal fake news detection system that combines BERT, DeBERTa, and ViT models. The BERT and DeBERTa models handle textual data to classify news as either real or fake, while the ViT model is trained separately to detect AI-generated vs real images, helping assess the authenticity of visual content.

Text Models — Fake News Detection

• Architecture: Ensemble of BERT-base and DeBERTa-base
• Task: Binary Text Classification (REAL vs FAKE)
• Training Framework: PyTorch using 🤗 Transformers
• License: MIT

Vision Model — AI-Generated Image Detection

• Architecture: Vision Transformer (ViT-base, vit-base-patch16-224)
• Task: Binary Image Classification ('REAL' vs 'AI-GENERATED')
• Training Framework: TensorFlow/Keras or PyTorch using Transformers
• License: MIT

Dataset

The dataset is a custom collection combining:

• News content (title + body)
• Labels: 0 = FAKE, 1 = REAL

❗Disclaimer

• This project is for educational and experimental purposes only.
• It is not suitable for real-world fact-checking or serious decision-making.
• The model uses a simple binary classifier and does not verify factual correctness.
• Model may sometimes misclassify text with unclear or misleading context, and images that are abstract, artistic, or difficult to distinguish from real content.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("fauxNeuz/BertAndDeberta")
model = AutoModelForSequenceClassification.from_pretrained("fauxNeuz/BertAndDeberta")

text = "Government confirms policy updates in healthcare sector."
inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)
pred = outputs.logits.argmax(dim=-1).item()
print("REAL" if pred == 1 else "FAKE")