AUTHEN AI Media Authenticity & Forensic Engine

AUTHEN Core is an evidence-centric, multimodal AI authenticity and forensic verification engine designed to detect synthetic, manipulated, and deepfake content across image, video, and audio modalities.

Built from first principles (governed by AGENTS.md), AUTHEN Core replaces naive binary classifiers with a calibrated evidential reasoning framework capable of generalizing to unseen generative AI models.

---

🏛️ System Architecture

AUTHEN Core operates on a four-tier forensic pipeline:

                            ┌─────────────────────────────────┐
                            │    Input Media (Img/Vid/Aud)    │
                            └────────────────┬────────────────┘
                                             │
                        ┌────────────────────▼────────────────────┐
                        │   Media Ingestion & Invariant Preproc   │
                        └────────────────────┬────────────────────┘
                                             │
     ┌───────────────────────────────────────┼───────────────────────────────────────┐
     │                                       │                                       │
┌────▼─────────────────┐            ┌────────▼─────────────┐            ┌────────────▼──────────┐
│ Image Forensic Core  │            │ Video Forensic Core  │            │  Audio Forensic Core  │
│ - DINOv2 / CLIP ViT  │            │ - Bidirectional SSM  │            │  - WavLM + LoRA SSL   │
│ - 2D Haar Wavelet    │            │   (Mamba-style O(T)) │            │  - Mel-Spectrogram    │
│ - SRM + Bayar Noise  │            │ - Temporal Inconsist.│            │  - F0 / Jitter/Shimmer│
│ - Local Patch Scan   │            │ - Frame Transitions  │            │  - Phase Anomalies    │
└────┬─────────────────┘            └────────┬─────────────┘            └────────────┬──────────┘
     │                                       │                                       │
     └───────────────────────────────────────┼───────────────────────────────────────┘
                                             │
                             ┌───────────────▼───────────────┐
                             │   Structured Forensic Trace   │
                             │     [ForensicEvidence, ...]   │
                             └───────────────┬───────────────┘
                                             │
                             ┌───────────────▼───────────────┐
                             │  Dempster-Shafer Fusion & EDL │
                             │  - Dirichlet Uncertainty (u)  │
                             │  - Conflict & Disagreement    │
                             └───────────────┬───────────────┘
                                             │
                             ┌───────────────▼───────────────┐
                             │    Authenticity Assessment    │
                             │  - Verdict (Auth/Synth/Manip) │
                             │  - Calibrated Confidence      │
                             │  - Epistemic Uncertainty (u)  │
                             │  - Traceable Evidence Chain   │
                             └───────────────────────────────┘

---

⚡ Performance Benchmarks (Intel Core CPU, Single-Threaded)

Subsystem	Parameters	Model Size	Batch-1 Latency (p50)	p95 Latency
Image Core (224x224)	4.35 M	17.41 MB	9.22 ms	10.47 ms
Audio Core (3s @ 16kHz)	2.46 M	9.85 MB	12.78 ms	20.31 ms
Video Core (16 frames)	2.41 M	9.65 MB	14.38 ms	19.95 ms
EDL Fusion Head	0.099 M	0.40 MB	0.18 ms	0.20 ms
Full Pipeline (E2E)	9.32 M	37.30 MB	11.05 ms	11.72 ms

• Expected Calibration Error (ECE): 0.0094 ($< 1%$)
• Selective Prediction: Automatic abstention (Verdict.INCONCLUSIVE) when evidence uncertainty $u > 0.50$.

---

🚀 Quickstart

Installation

# Clone the repository
git clone https://github.com/MxHabob/authen-ml.git
cd authen-ml

# Install dependencies in development mode
pip install -e .[dev]

Python API Usage

import torch
from authen_core import (
    AuthenPipeline,
    ImageForensicAnalyzer,
    AudioForensicAnalyzer,
    VideoForensicAnalyzer,
    EvidenceAggregator,
)

# 1. Initialize Pipeline with analyzers
pipeline = AuthenPipeline(
    analyzers=[
        ImageForensicAnalyzer(),
        AudioForensicAnalyzer(),
        VideoForensicAnalyzer(),
    ],
    aggregator=EvidenceAggregator(abstention_threshold=0.50),
)

# 2. Run Forensic Analysis on an image
dummy_image = torch.randn(3, 224, 224)
assessment = pipeline.analyze({"image": dummy_image}, media_id="sample_001")

# 3. Inspect structured verdict and calibrated evidence
print(f"Verdict:     {assessment.verdict.value}")
print(f"Confidence:  {assessment.confidence:.4f}")
print(f"Uncertainty: {assessment.uncertainty:.4f}")
print(f"Evidence Count: {len(assessment.evidence_chain)}")

for ev in assessment.evidence_chain:
    print(f" - [{ev.modality.value}] {ev.analyzer}: {ev.description} (conf={ev.calibrated_confidence:.2f})")

---

🧪 Running Tests & Benchmarks

# Run the complete test suite (57 unit & integration tests)
python -m pytest tests/test_authen_core.py -v

# Run the performance and calibration benchmark
python scripts/benchmark_authen_core.py --device cpu --num-runs 30

---

📚 Scientific & Engineering Documentation

• State of the Art Research (2024–2026)
• Threat Model & Anti-Forensics
• Algorithm Design Specification
• Evidence Schema Specification
• Benchmark Protocol
• Ablation Study Plan
• Efficiency & Optimization Plan
• Reproducibility & Experiment Guide
• System Specification Document
• ML Model Card
• Final Algorithmic Decision Document