| --- |
| title: VeriDex |
| emoji: 🛡️ |
| colorFrom: blue |
| colorTo: indigo |
| sdk: gradio |
| sdk_version: 4.19.2 |
| app_file: app.py |
| pinned: false |
| short_description: Multimodal Claim Verification Engine |
| --- |
| |
| # VeriDex: A Fake News Detection System Using Text Classification and Image Forensics |
|
|
| **Author:** Rakesh Kumar Raut |
|
|
|
|
| --- |
|
|
| ## 📖 Abstract & Overview |
|
|
| AI-generated misinformation has exploded, and automated fact-verification systems are struggling to keep up. Modern Large Language Models (LLMs) churn out text that is almost impossible to tell apart from real language, while diffusion-based image synthesis creates deepfakes so realistic they make false claims look credible. These combined threats create a crucial "Zero-Day window" right after a big event, where independent fact-checks are missing, and evidence-based systems are helpless. |
|
|
| **VeriDex (Verified Deception Index)** is a Hybrid Multi-Modal Credibility Assessment System designed to tackle text from LLMs, AI-generated images, and the Zero-Day problem in one principled and transparent package. |
|
|
| VeriDex runs three independently trained deep-learning pipelines and brings their results together through a mathematically derived **Hybrid Resolution Engine**. Tested on a custom 500-claim adversarial benchmark, VeriDex achieves state-of-the-art results, significantly outperforming text-only and retrieval-only baselines. |
|
|
| --- |
|
|
| ## 🏗️ Core Architecture (The Three Pipelines) |
|
|
| VeriDex takes a modular approach, with three parallel pipelines feeding into a central Hybrid Resolution Engine. This complementary redundancy ensures that each pipeline thrives exactly where the others fall short. |
|
|
| ### 1. Pipeline A: Linguistic Deception Analyzer (HierFND) |
| Catches linguistic deception based on writing style and text artifacts. |
| * **Model:** Fine-tuned RoBERTa-base sequence classifier. |
| * **Training Data:** LIAR, ISOT, and WELFake datasets. |
| * **Output:** A Linguistic Risk Score ($R_{ling}$) reflecting the probability of the text being machine-generated or stylistically deceptive. |
| |
| ### 2. Pipeline B: Retrieval-Augmented Stance Aggregator (StanceFormer) |
| Verifies claims against actual evidence via real-time web retrieval. |
| * **Model:** Fine-tuned DeBERTa-v3-base Natural Language Inference (NLI) model. |
| * **Mechanism:** The claim triggers a web search API, pulling the top $k=10$ articles. |
| * **Domain-Credibility RAG:** Each article’s domain gets mapped to a Media Bias/Fact Check (MBFC) credibility multiplier (1.5 for high-credibility, 1.0 for neutral, 0.2 for questionable/satire). |
| * **Output:** An Evidence Consensus Score ($E_{pro}$) indicating the factual support for the claim. |
|
|
| ### 3. Pipeline C: AI Image Forensics (CRAFT) |
| Pinpoints AI-generated images (deepfakes). |
| * **Architecture:** Combines a CLIP ViT-B/32 semantic branch with LoRA adapters and a 2D FFT-based **FrequencyBranch**. |
| * **Mechanism:** Merges 512-dimensional semantic features with 128-dimensional spectral features to catch checkerboard artifacts from GANs and diffusion models. |
| * **Output:** Image forensic classification (AI-Generated vs. Authentic) powered by a two-layer MLP head. |
|
|
| --- |
|
|
| ## ⚙️ Hybrid Resolution Engine |
|
|
| The core blend mixes linguistic risk with evidence-based risk using an empirically derived blending weight ($\alpha^* = 0.40$): |
|
|
| $$R_{total} = (R_{ling} \times 0.40) + ((1 - E_{pro}) \times 0.60)$$ |
| |
| This 40/60 split keeps pure text classifiers from missing LLM-generated fakes, while also countering the failures of retrieval-only approaches during the Zero-Day window. |
| |
| ### Fail-Safe Override Mechanisms |
| Two formal override rules ensure system robustness: |
| 1. **Explicit Debunk Override:** If a highly credible article (weight $\ge 1.4$) strongly refutes the claim and uses terms like `"fact check"`, `"debunked"`, or `"false"`, the system forces $R_{total} \leftarrow \max(R_{total}, 0.90)$. This stops a clever lie from outweighing legitimate fact-checking. |
| 2. **Image Forensics Override:** If an image is flagged as AI-Generated with $\ge 85\%$ confidence, the system forces $R_{total} \leftarrow \max(R_{total}, 0.85)$. |
| |
| --- |
| |
| ## 📊 Empirical Results & Benchmarks |
| |
| VeriDex was rigorously evaluated on a custom **500-claim adversarial benchmark** encompassing Standard Fake News, Standard Real News, Zero-Day Claims, and LLM-Generated Fakes. |
| |
| ### Integrated System Performance |
| | Metric | Score | Category Accuracy Breakdown | |
| | :--- | :--- | :--- | |
| | **Accuracy** | 94.80% | Standard Fake: 97.5% | |
| | **AUC-ROC** | 0.965 | Standard Real: 98.0% | |
| | **Precision (Fake)** | 0.9751 | Zero-Day: 89.3% | |
| | **Recall (Fake)** | 0.9352 | LLM-Generated: 86.7% | |
| | **Macro-F1** | 0.9468 | | |
| |
| > **Comparison to State-of-the-Art (SOTA):** VeriDex beats FakeBERT by +4.8 pp, VeraCT Scan by +3.3 pp, and classical TF-IDF+SVM approaches by +12.8 pp. |
| |
| ### Image Forensics (CRAFT) Performance |
| Tested on a rigorous cross-generator protocol against 50,000 images (DALL-E 2, Midjourney v5, ProGAN, StyleGAN). |
| * **In-Distribution Accuracy:** 91.5% |
| * **Cross-Generator Accuracy:** 84.3% (Outperforming GenDet CVPR 2024 by 4.9 pp) |
| * **AUC-ROC:** 0.921 |
| |
| --- |
| |
| ## 🌟 Novel Contributions |
| |
| 1. **Empirical 40/60 Weighting:** The first systematic grid search (101 configurations) over blending weights for linguistic and evidence information, backed by a bootstrapped 95% confidence interval [0.340, 0.462]. |
| 2. **Domain-Credibility RAG:** MBFC credibility multipliers act as continuous weights in evidence aggregation, boosting accuracy by +1.8 pp and blocking echo-chamber manipulation. |
| 3. **Explicit Debunk Override:** A formal fail-safe ensuring high-credibility explicit fact-checks cannot be diluted by text model confidence (+1.4 pp accuracy). |
| 4. **CLIP + LoRA + FrequencyBranch:** A novel parameter-efficient deepfake detector that slashes generalization degradation by 32% compared to CLIP-only setups. |
| |
| --- |
| |
| ## 🚀 Setup & Installation |
| |
| ### Prerequisites |
| - Python 3.9+ |
| - CUDA-enabled GPU (Highly Recommended for inference speed) |
| |
| ### 1. Clone & Install |
| ```bash |
| git clone https://github.com/RakeshRautDev/VeriDex.git |
| cd VeriDex |
| pip install -r requirements.txt |
| ``` |
| |
| ### 2. Environment Variables |
| Copy `.env.example` to `.env` inside `VeriDex_WebApp` and configure your API keys. |
| ```bash |
| cd VeriDex_WebApp |
| cp .env.example .env |
| ``` |
| Edit `.env` to include your search API keys: `TAVILY_API_KEY=your_key_here` |
|
|
| ### 3. Model Weights Setup |
| > [!IMPORTANT] |
| > Due to GitHub's file size constraints (100MB max per file), heavy binary model weights are **NOT included** in this repository. |
| > |
| > You must download the pre-trained weights separately and place them in their respective model directories: |
| > - `models/fakeNewsModel/pytorch_model.bin` |
| > - `models/stanceModel/model.safetensors` |
| > - `models/imageDetectionModel/best_model.pth` |
|
|
| ### 4. Run the Web Application |
| ```bash |
| cd VeriDex_WebApp |
| uvicorn app:app --host 0.0.0.0 --port 8000 --reload |
| ``` |
| Open `http://localhost:8000/static/index.html` in your browser to access the verification dashboard. |
|
|
| --- |
|
|
| ## 📝 License |
| MIT License |
|
|
