Naei777/menaml-2026-digital-integrity-trackA

MenaML 2026 - Digital Integrity Challenge

Track A: Social Media & Influencer Authenticity

Team Information

• —Lead: Naeimeh Najafizadeh Sari
• —Contributor: Hadi Jahanshahi

System Overview

This is a dual-module AI image authenticity detection system that combines:

1. 1.Module 1: Forensic Signal Detector (ViT + FFT frequency analysis)
2. 2.Module 2: Vision-Language Model reasoning (semantic analysis with LLaVA 7B)

Installation

Prerequisites

• —Python 3.8+
• —macOS or Linux
• —Ollama installed (for VLM module)

Setup Instructions

1. 1.Install Ollama (if not already installed):

# macOS
brew install ollama

# OR download from https://ollama.com/download

# Start Ollama and pull the LLaVA model
ollama pull llava:7b

1. 1.Install Python dependencies:

pip install -r requirements.txt

Usage

Single Image Analysis

python predict.py --image path/to/image.jpg

Batch Processing

python predict.py --input_dir path/to/images/ --output_file predictions.json

Output Format

The system generates JSON output with:

• —authenticity_score: 0.0-1.0 (0.0 = authentic, 1.0 = manipulated)
• —manipulation_type: Classification of manipulation type
• —vlm_reasoning: Natural language explanation

Architecture

• —Forensic Detector: buildborderless/CommunityForensics-DeepfakeDet-ViT
• —Frequency Analysis: FFT for detecting GAN/diffusion artifacts
• —VLM Reasoner: LLaVA 7B via Ollama
• —Fusion Strategy: Weighted combination (65% forensic, 35% VLM)

Performance

Tested on balanced dataset (10 AI-generated, 10 real images):

• —AI Detection: 100% (all synthetic identities caught)
• —Real Detection: 90% automatic pass, 10% human review
• —Overall Accuracy: 95% definitive decisions
• —Zero false positives, zero false negatives

Technical Report

See technical_report.pdf for detailed methodology, model comparison, and analysis.

Model Weights

This system uses pre-trained models that are downloaded automatically:

1. 1.Forensic Detector: Downloaded via HuggingFace transformers library
2. 2.VLM Analyzer: Requires Ollama with llava:7b model (see installation steps above)