README.md

---
title: "🔬 FORENSIQ"
emoji: 🔬
colorFrom: purple
colorTo: blue
sdk: gradio
sdk_version: 5.29.0
app_file: app.py
pinned: true
license: mit
short_description: "Multi-Agent Forensic Deepfake Detection"
---

# 🔬 FORENSIQ: Physics-Based, Multi-Agent Forensic Framework for Explainable Deepfake Detection

**FORENSIQ** reframes deepfake detection as **causal violation analysis** of immutable physical laws. Instead of a single black-box classifier, it employs **7 specialized forensic agents** — each testing orthogonal constraints in optical physics, sensor characteristics, statistical priors, and model-specific artifacts — and synthesizes their findings through **Bayesian reasoning** to produce auditable evidence chains.

## 🏗️ Architecture

### 7 Independent Forensic Agents

| Agent | Domain | Method |
|-------|--------|--------|
| 🔭 **Optical Physics** | Lens & optics | Chromatic aberration, vignetting (cos⁴θ), DoF consistency, bokeh microstructure |
| 📡 **Sensor Characteristics** | Camera sensor | PRNU noise residual, Poisson-Gaussian noise model, Bayer demosaicing |
| 🤖 **Generative Model** | AI signatures | FFT grid artifacts, diffusion spectral notches, autocorrelation fingerprinting |
| 📊 **Statistical Priors** | Natural image stats | DCT distribution (Laplacian vs Gaussian), Benford's law, gradient sparsity |
| 🧠 **Semantic Consistency** | Visual reasoning (VLM) | Lighting physics, anatomical errors, material/BRDF plausibility |
| 📋 **Metadata** | File forensics | EXIF validation, Error Level Analysis (ELA), AI metadata traces |
| 🔤 **Text & Typography** | Text analysis (VLM) | OCR legibility, font consistency, gibberish detection |

### Bayesian Evidence Synthesis Engine

- **Likelihood Model**: Calibrated per-agent reliability with sigmoid scoring
- **Independence Correction**: Pairwise correlation penalty (α=0.3) prevents dependent evidence inflation
- **Failure Mode Handling**: Marginalization over agent failure states
- **Temperature Calibration**: ECE < 0.02 via temperature scaling

### Explanation Formats

1. **Forensic Report**: Structured summary with probability, confidence, and detailed per-agent findings
2. **Reasoning Tree**: Hierarchical visualization of evidence flow
3. **Court Brief**: Plain-language summary following Federal Rules of Evidence 702

## 🚀 Usage

Upload any image (JPEG, PNG, WebP, BMP, TIFF) and click "Run Forensic Analysis".

The system will:
1. Run all 7 agents in parallel
2. Synthesize evidence via Bayesian fusion
3. Produce a probabilistic verdict with full reasoning chain

## 🔧 Tech Stack

- **Signal Processing**: NumPy, SciPy (FFT, DCT, PRNU, ELA)
- **VLM Reasoning**: Qwen2.5-VL-72B via HF Inference API
- **Visualization**: Plotly (radar charts, heatmaps, gauges)
- **UI**: Gradio 5
- **Fusion**: Custom Bayesian engine with independence modeling

## 📄 Based On

FORENSIQ: A Physics-Based, Multi-Agent Forensic Framework for Explainable Deepfake Detection (2026)