Fake News Detector

A multimodal fake news detection web app built with Streamlit, powered by DistilBERT, CLIP, Whisper, and audio prosody features. Detects misinformation in both text articles and videos.

Features

• Article Analysis — Paste text or provide a URL; the app scrapes and classifies the content using a DistilBERT-based model
• Video Analysis — Upload a video or provide a URL; the app transcribes speech (Whisper), extracts visual frame embeddings (CLIP), and audio prosody features (librosa) for multimodal classification
• Uncertainty flagging — Predictions below 70% confidence are flagged as ⚠️ UNCERTAIN instead of a hard verdict
• Word cloud — Visual breakdown of the most prominent words in analyzed text

Project Structure

fake-news-detector/
├── app.py              # Entry point — page config and tab wiring
├── config.py           # Constants (DEVICE, MAX_LEN, thresholds) and sample articles
├── loaders.py          # Cached model and tokenizer loading (@st.cache_resource)
├── scraper.py          # URL scraping via newspaper3k and BeautifulSoup
├── predictor.py        # Text cleaning, tokenization, and inference functions
├── transcriber.py      # Whisper-based audio transcription (file and URL)
├── ui.py               # All Streamlit UI rendering (tabs, verdicts, word clouds)
├── model.py            # TextClassifier and MultimodalClassifier model definitions
├── features.py         # CLIP visual and librosa audio feature extraction
├── train_text.py       # Training pipeline for the text model (WELFake + ISOT)
├── train.py            # Training pipeline for the video/multimodal model
├── prep_mklab.py       # MKLab FVC dataset video downloader
├── requirements.txt    # Python dependencies
├── data/               # (User-created) Holds input datasets and cached features
└── model/              # (User-created) Holds trained model checkpoints

Models

Text Model — TextClassifier

• Backbone: DistilBERT (distilbert-base-uncased), fully frozen
• Head: MLP (Linear → LayerNorm → ReLU → Dropout → Linear)
• Input: Mean-pooled 768-dim token embeddings (max 256 tokens)
• Training data: Up to 20,000 balanced samples from WELFake (and optionally ISOT)
• Optimization: Embeddings pre-computed once and cached to data/text_embedding_cache.npz; only the MLP head is trained on subsequent runs

Video Model — MultimodalClassifier

• Text stream: BERT (bert-base-uncased, last 2 layers unfrozen) → 768-dim
• Visual stream: CLIP ViT-B/32, 8 frames averaged → 512-dim → 256-dim projection
• Audio stream: 13 prosody features (F0, spectral centroid, RMS, 7 MFCCs) → 128-dim projection
• Fusion: Cross-modal attention (text ↔ visual) → concatenation → classifier head
• Training: 25 epochs, AdamW + cosine LR decay, batch size 4, 25% modality dropout

Confidence Threshold

Verdicts with confidence below 70% are displayed as ⚠️ UNCERTAIN to avoid misleading low-confidence classifications. This threshold is configured in config.py via CONFIDENCE_THRESHOLD.

Requirements

• streamlit — web app framework
• transformers — DistilBERT and BERT models
• openai-whisper — speech transcription
• open-clip-torch — CLIP visual embeddings
• librosa — audio feature extraction
• newspaper3k — article scraping
• yt-dlp — video downloading
• wordcloud, matplotlib — visualization

Built With

• clip
• librosa
• natural-language-processing
• python
• pytorch
• streamlit
• web-scraping
• whisper