README.md

metadata

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)