🛡️ TrustCheck OSINT Platform

MenaCraft Hackathon 2026 Submission Live Multi-Axis Forensic Verification Suite

TrustCheck is a high-performance, expert-level media verification platform designed to combat the "liar's dividend" and the rise of high-fidelity deepfakes. It moves beyond single-prompt AI checks by running a parallel, multi-axis investigation pipeline that resolves contradictions in forensic data.

🏗️ System Architecture: The 5-Axis Forensic Engine

TrustCheck doesn't just ask if something is "real." It investigates the content across 5 independent dimensions:

1. Visual Forensics (Axis A)

• Neural Discovery: Leverages LLaMA 3.2 Vision and Qwen2.5-VL to detect AI generation fingerprints, lighting inconsistencies, and neural artifacts.
• Latent Manifold Reconstruction (Math): A dual-method mathematical detector that analyzes the "neural footprint" of images:
  • Method 1: HF Classifier: State-of-the-art inference (Organika/sdxl-detector) with ~99% accuracy across all AI generators.
  • Method 2: VAE Manifold Analysis: Uses a Variational Autoencoder to measure PSNR/MSE and KL Divergence. AI images lie perfectly on the generative latent prior, while real photos deviate significantly.
• Local ELA: Analyzes JPEG re-save artifacts to identify exactly which regions of an image have been digitally manipulated.

2. Acoustic Forensics (Axis B)

• Neural Voice Detection: Analyzes audio for "Robotic Perfection."
• Micro-perturbation Analysis: Detects Jitter (F0 instability) and Shimmer (amplitude instability) that are present in human speech but often missing in Neural TTS (Canva, ElevenLabs).
• Vocoder Fingerprinting: Identifies spectral-contrast anomalies characteristic of digital vocoders and neural synthesis.

3. Contextual Consistency (Axis C)

• Semantic Mapping: Checks if the visual content matches the user claim (e.g., "Is that really a forest in Tunisia?").
• Fact-Check Integration: Real-time cross-referencing with the Google Fact Check Tools API to identify known viral misinformation.

4. Digital Footprint (Axis D)

• EXIF Forensics: Extracts deep metadata from files including camera serials, lens focal length, and software history (Adobe Photoshop/Canva tags).
• GPS-Stripping Detection: Flags "high-suspicion" signals when GPS hardware exists but coordinates have been deliberately scrubbed.

5. Source Credibility (Axis E)

• Deterministic Signals: Checks for suspicious TLDs (.xyz, .top), typosquatting, and emotional amplification patterns.
• Reputational Assessment: Blends automated domain forensics with LLM reputational analysis.

---

🧠 Master Forensic Synthesis

The platform features a Chief Investigator reasoning layer (LLaMA 3.3 70B). This layer reviews all 5 axis reports, identifies contradictions (e.g., "AI-flagged audio that appears perfectly studio-clean"), and produces a final Weighted Risk Verdict (CRITICAL to LOW).

🛠️ Technical Stack

• Backend: Python (FastAPI), ThreadPoolExecutor (Parallel analysis).
• ML/LLM: Groq (LLaMA Series), OpenRouter (Vision Models), DeepSeek, HuggingFace Hub (Classifiers).
• Forensics/Math: PyTorch (VAE Latent Manifold), Librosa (Acoustics), OpenCV (Image ELA), piexif.
• Database: Supabase (Forensic history + JSONB).
• Frontend: Next.js 15+ "Investigation Room" Dashboard. A high-performance, reactive React frontend using Radix UI and Tailwind CSS for a premium "Cyber-Forensic" aesthetic. Built with static-export for seamless FastAPI integration.

🚀 Getting Started

1. Clone the repository.
2. Install dependencies: pip install -r requirements.txt
3. Configure .env with your API keys (Groq, OpenRouter, Google, Supabase).
4. Run the engine: python -m uvicorn app:app --reload
5. Access the command center at http://localhost:8000/ui

---

Developed for the MenaCraft Hackathon — 2026