A-IQ: AI Image Detection for macOS

A powerful, privacy-focused macOS application that detects AI-generated images using advanced machine learning and forensic analysis. A-IQ combines five independent detection methods to provide reliable results — all while processing images entirely on your device.

100% Local Processing — Your Photos Never Leave Your Mac

Unlike web-based AI detectors that upload your images to remote servers, A-IQ runs entirely on your device. Your photos are analyzed locally using your Mac's Neural Engine — nothing is ever sent to the cloud.

• ✓ No account required
• ✓ No internet connection needed
• ✓ No data collection or analytics
• ✓ No privacy concerns

Your images stay on your Mac. Period.

Quick Start

Simple Workflow

1. Drag any image onto A-IQ
2. Drop it in the analysis window
3. Get an instant AI probability score

Analysis typically completes in under 3 seconds.

Multiple Input Methods

Method	How
Drag and Drop	Drag any image onto the A-IQ window
Open File	Click "Open File" or press Cmd+O
Paste	Copy an image and press Cmd+V
Batch Analysis	Select "Open Folder" to analyze multiple images

Multi-Signal Detection

A-IQ doesn't rely on a single method — it combines five independent detection signals for maximum accuracy:

Signal	Default Weight	With Faces	Method
ML Detection	40%	35%	SigLIP Vision Transformer neural network
Provenance	30%	25%	C2PA content credentials verification
Metadata	15%	10%	EXIF/IPTC anomaly analysis
Forensics	15%	10%	Error Level Analysis (ELA) + FFT
Face-Swap	—	20%	Deepfake detection (Vision + ML)

When faces are detected, weights automatically redistribute to include face-swap analysis.

Understanding Results

A-IQ provides a confidence score from 0-100%:

Score	Classification	Meaning
< 30%	Likely Authentic	Low probability of AI generation
30-70%	Uncertain	Review recommended
> 70%	Likely AI-Generated	High probability of AI generation
100%	Confirmed AI-Generated	C2PA credentials prove AI origin

Each analysis includes:

• Overall AI Probability Score (0-100%)
• Classification: Likely Authentic, Uncertain, or Likely AI-Generated
• Signal Breakdown: See how each detection method contributed
• Evidence Summary: Specific indicators found in the image
• Visual Forensics: View ELA overlays to see detected anomalies

Analysis Examples

Signal Breakdown	Image Metadata	Error Level Analysis

Face-swap Detection	Evidence Summary

How Detection Works

1. Neural Network Analysis (40% weight)

A-IQ uses SigLIP (Sigmoid Loss Image-Language Pretraining), a state-of-the-art vision transformer. This advanced neural network recognizes subtle patterns that distinguish AI-generated content from authentic photographs.

Technical Details:

• Architecture: Vision Transformer (ViT)
• Input: 224×224 pixels (center crop)
• Output: [AI probability, Human probability]
• Hardware: Neural Engine on Apple Silicon, GPU fallback on Intel
• Timeout: 2.0 seconds
• Detects content from Grok, DALL-E, Midjourney, Stable Diffusion, Flux, Adobe Firefly, and 60+ other AI generators

2. Content Credentials (C2PA) Verification (30% weight)

A-IQ performs deep C2PA analysis using the industry-standard Content Authenticity Initiative protocol backed by Adobe, Microsoft, Google, and the BBC.

Core Verification:

• Cryptographic signature validation against trusted signer database
• Detection of 60+ AI tools in credential claims
• Definitive proof when valid credentials indicate AI origin

Enhanced C2PA Parsing:

• AI Generation Assertions: Parses c2pa.ai_generative_info and c2pa.synthetic for explicit AI disclosure
• Model/Prompt Extraction: Extracts AI model name, generation prompt, and parameters (cfg_scale, steps, seed)
• Ingredient Analysis: Detects AI-generated parent images in composites via c2pa.ingredients
• Provenance Chain Scanning: Identifies AI tool usage at any point in the image's history
• Training Status: Checks c2pa.ai_training assertions

Nuanced Scoring:

Score	Condition
100%	Valid C2PA + known AI tool or explicit AI assertion
95%	Explicit AI assertion (without valid credentials)
85%	AI tool in provenance chain
80%	Parent image is AI-generated
70%	Tampered credentials
50%	No credentials (neutral)
20%	Valid non-AI credentials (likely authentic)

3. Metadata Forensics (15% weight)

Deep inspection across multiple metadata layers:

AI Software Detection (60+ signatures):

• Grok, DALL-E, Midjourney, Stable Diffusion, Flux, Fooocus, ComfyUI, and many more
• AI pipeline software (Pillow, PyTorch, OpenCV, ImageMagick)

Advanced Anomaly Detection:

• Missing EXIF: No EXIF data in JPEG files (common in AI-generated images)
• Timestamp Anomalies: Future dates, pre-digital era dates, inconsistencies
• Thumbnail Mismatch: Compares embedded thumbnail to main image (>25% difference flagged)
• Color Profile Analysis: Detects generic/uncalibrated profiles, camera brand vs profile mismatches
• JPEG Compression Analysis: Identifies AI pipeline software, unusual JFIF density
• XMP/IPTC Deep Inspection: Scans for AI terms (prompts, cfg scale, sampler, seed, etc.)

4. Forensic Image Analysis (15% weight)

Combines traditional and frequency domain analysis:

Error Level Analysis (ELA) — Traditional Forensics (60% of forensic score)

• Recompresses image at 90% quality and computes pixel-by-pixel differences
• AI-generated images often show uniform compression patterns
• Edited photos show localized differences at manipulation boundaries

FFT Frequency Spectrum Analysis — Advanced Forensics (40% of forensic score)

• 2D Fast Fourier Transform via Accelerate vDSP for hardware acceleration
• Radial power spectrum analysis (azimuthal averaging)
• Power law fitting: natural images follow 1/f^n decay (slope typically -2 to -3)
• Spectral peak detection for GAN upsampling artifacts
• Detects anomalous frequency patterns that betray synthetic origin

Constraints:

• Max resolution: 3840×2160 (downscaled if larger)
• FFT size: Power of 2, max 512×512
• 3-second timeout

5. Face-Swap & Deepfake Detection (20% weight when faces present)

When A-IQ detects faces, it activates a dedicated deepfake detection neural network trained on the FaceForensics++ dataset (~94% accuracy).

ML-Powered Detection:

• Uses SigLIP Vision Transformer fine-tuned for face manipulation detection
• Each face extracted with 20% padding for context
• Resized to 224×224 and analyzed by DeepfakeDetector model
• Per-face probability scores with visual bounding boxes

Dynamic Weight Behavior:

• When faces detected: Adds 20% weight, other detectors reduce proportionally
• When no faces: Returns neutral result, original weights apply

Per-Face Results:

• Deepfake probability score (0-100%)
• Classification: Low (<30%), Medium (30-70%), High (>70%)
• Visual bounding box highlighting analyzed regions
• List of detected artifacts with severity

Use Cases

• Journalists & Fact-checkers: Verify image authenticity before publishing
• Content Moderators: Screen submissions for AI-generated content
• Photographers: Prove your work is authentic
• Researchers: Analyze datasets for synthetic content
• Security Professionals: Investigate potential deepfakes
• Anyone: Satisfy curiosity about suspicious images

Batch Processing

• Folder scanning: Analyze entire folders at once
• Batch queue: Process multiple images in parallel (max 4 concurrent)
• History: Review all past analyses with search and filtering
• Export Reports: Generate PDF or JSON reports

Supported Formats

• JPEG / JPG
• PNG
• HEIC / HEIF
• WebP
• TIFF
• AVIF

Key Features

• Privacy-First: All processing happens locally. No images uploaded to any server.
• Offline Capable: Works without an internet connection.
• Native Performance: Optimized for Apple Silicon with Neural Engine acceleration.
• Batch Analysis: Analyze entire folders of images.
• Export Reports: Generate PDF or JSON reports of findings.
• History: Browse and search past analysis results (max 1000 records).
• Dark Mode: Full support for macOS appearance modes.
• Keyboard Shortcuts: Power user workflows with keyboard shortcuts.

Recent Improvements

v1.2 (January 2026)

New Features:

1. Grok/xAI Detection: Added comprehensive detection for xAI's Grok image generator
   • Detects: Grok, xAI, Aurora, Grok Imagine, Grok 2, grok-2-image
   • Note: xAI does not implement C2PA credentials; detection relies on ML analysis and metadata signatures

v1.1 (December 2025)

Critical Fixes:

1. Graceful Error Handling: Replaced fatal errors with user-friendly dialogs
2. Complete Menu Implementation: All menu commands now functional
3. Proper Task Cancellation: Analysis can be cancelled without errors
4. Comprehensive Error Handling: All errors are caught and displayed appropriately
5. Input Validation: File size limits (100MB) and security improvements
6. Constants Extraction: Centralized configuration for easier maintenance

Code Quality:

• Centralized constants in AnalysisConstants.swift
• Improved error handling patterns
• Better security with path validation and symlink resolution
• Enhanced user feedback
• Removed dead code (unused variables in MetadataAnalyzer)
• Added helper methods for consistent optional handling in ResultAggregator
• Standardized use of AnalysisConstants.neutralScore for default values

Requirements

• macOS 14.0 (Sonoma) or later
• Apple Silicon (M1/M2/M3/M4) recommended for best performance
• Intel Macs supported
• 400MB disk space
• No internet connection required

Installation

For End Users

1. Download A-IQ.dmg from the Releases page
2. Open the DMG and drag A-IQ.app to your Applications folder
3. Right-click A-IQ.app → "Open" on first launch (macOS security)
4. Launch normally afterward

Note: The DMG contains all required models and resources. No additional downloads needed.

Architecture

A-IQ is built as a native macOS app using SwiftUI and modern concurrency patterns:

A-IQ/
├── App/           # Application entry point and state
├── Analysis/      # Orchestration and result aggregation
├── Detectors/     # ML, Provenance, Metadata, Forensic, FaceSwap analyzers
├── Models/        # Data structures and protocols
├── Views/         # SwiftUI interface
├── Input/         # Image acquisition (file, clipboard, drag-drop)
├── Storage/       # SwiftData persistence
├── Export/        # PDF/JSON report generation
├── Settings/      # User preferences
└── Resources/     # ML model, c2patool, trust list

Technical Highlights

Concurrency Model:

• Actor-based architecture for thread safety
• Parallel detector execution via async let
• Max 4 concurrent analyses with memory throttling (2GB limit)
• Cancellable analysis tasks with proper cleanup

Performance:

• Neural Engine acceleration on Apple Silicon
• GPU fallback on Intel Macs
• Large images downscaled to 4K for forensic analysis
• Lazy model loading to reduce startup time

Storage:

• SwiftData for analysis history
• Max 1000 records with automatic cleanup
• JPEG-compressed thumbnails (70% quality)
• Full results stored as JSON

Security:

• Path validation with symlink resolution prevents path traversal
• File size limits (100MB maximum)
• Input sanitization for all file paths
• Sandboxed external tool execution (c2patool)

Building from Source

For Developers

Important: The repository does not include large binary files (ML models, c2patool). You must download these separately before building.

Prerequisites

• Xcode 15.0 or later
• macOS 14.0 SDK
• Git

Step 1: Clone Repository

git clone https://github.com/hmohamed01/A-IQ.git
cd A-IQ

Step 2: Download Required Resources

The following files are required but not included in the repository (too large for GitHub):

1. ML Model (AIDetector.mlmodelc) - 164MB

   • SigLIP Vision Transformer for AI detection
   • Place in A-IQ/Resources/AIDetector.mlmodelc

2. Deepfake Model (DeepfakeDetector.mlmodelc) - 164MB

   • SigLIP Vision Transformer for face-swap detection
   • Source: prithivMLmods/deepfake-detector-model-v1
   • Place in A-IQ/Resources/DeepfakeDetector.mlmodelc

3. c2patool binary - 34MB

   • Download from C2PA releases
   • Place in A-IQ/Resources/c2patool
   • Set executable: chmod +x A-IQ/Resources/c2patool

4. Trust List (trust_list.json)

   • JSON file listing trusted C2PA signers
   • Place in A-IQ/Resources/trust_list.json

Step 3: Build the App

Command Line:

xcodebuild build -scheme A-IQ -configuration Release -destination 'platform=macOS'

Xcode:

open A-IQ.xcodeproj
# Then Product → Build (⌘B)

The built app will be in build/Build/Products/Release/A-IQ.app

Step 4: Create DMG Installer (Optional)

bash scripts/package.sh

This creates dist/A-IQ.dmg with styled background and Applications symlink.

Configuration

Settings

Access via A-IQ → Settings or ⌘,:

• Sensitivity Threshold: Adjust detection sensitivity (0.0-1.0)
• Export Format: Default format for reports (PDF/JSON)
• Auto-Analyze on Drop: Automatically analyze dropped images
• Store Thumbnails: Enable/disable thumbnail storage in history
• History Retention: Days to keep history (0 = forever)

Constants

Key configuration values are centralized in Models/AnalysisConstants.swift:

• maxConcurrentAnalyses: Maximum parallel analyses (default: 4)
• memoryThresholdBytes: Memory limit for throttling (default: 2GB)
• likelyAuthenticThreshold: Score threshold for authentic (default: 0.30)
• likelyAIGeneratedThreshold: Score threshold for AI (default: 0.70)

Privacy Promise

A-IQ is built on a simple principle: your photos are yours.

• No analytics or telemetry
• No network requests
• No data collection
• Works completely offline after installation
• Full macOS app sandbox with minimal permissions

In an age of data harvesting and privacy erosion, A-IQ lets you detect AI images without sacrificing your privacy.

License

MIT License - See LICENSE file for details.

Acknowledgments

• C2PA for the content provenance standard
• Content Authenticity Initiative for c2patool
• SigLIP research for the vision transformer architecture
• Research community for advances in synthetic media detection

---

Version: 1.3 Last Updated: January 2026