R&D

Our Technology

Proprietary Models - Exceptional Accuracy - Unbreakable Security

At ARMA, we built our models from scratch. We did not rely on off-the-shelf or open-source solutions. We invested years of R&D to achieve 99.9% face recognition accuracy, outperforming most commercial solutions available.

Our models are trained on millions of faces from diverse ethnicities and conditions, making them the most diverse and accurate in the MENA region.

Model Architecture

State-of-the-Art Deep Learning & Neural Networks

The ARMA model is built on a hybrid architecture combining:

  • Convolutional Neural Networks (CNN) - for deep facial feature extraction
  • Transformer Architecture - for analyzing relationships between facial landmarks
  • ArcFace Loss Function - for improved discrimination between similar faces
  • Knowledge Distillation - for model compression while maintaining accuracy

Our Models & Technologies

Six integrated models built from scratch delivering unmatched performance

Face Recognition Model
Most Accurate

Face Recognition Model

Our proprietary face recognition model with 99.9% accuracy - the most advanced in the Middle East

The ARMA face recognition model is the result of 5 years of intensive research. Trained on over 10 million faces across diverse ethnicities, ages, and lighting conditions.

It uses advanced Deep Face Embedding algorithms that convert each face into a unique 512-dimensional digital fingerprint (Face Embedding), enabling precise and ultra-fast matching.

The model recognizes a face in under 0.1 seconds even in databases containing millions of faces.

Accuracy: 99.9%
Speed: < 0.1s
Technical Specifications
  • Dimensions: 512-dimension Face Embedding
  • Size: Only 6.8MB (lightweight for edge devices)
  • Training: 10M+ faces | 50+ ethnicities | 100+ countries
  • Framework: PyTorch + ONNX Runtime
  • Acceleration: NVIDIA TensorRT | Intel OpenVINO | ARM NN
Liveness Detection Model
Anti-Spoofing

Liveness Detection Model

Advanced 3D anti-spoofing system to detect fraud and forgery attempts

The ARMA liveness detection model is the first line of defense against fraud. It distinguishes real faces from forgery attempts including:

  • Printed photos - even high-resolution ones
  • Digital videos - from screens and devices
  • 3D masks - silicone and latex
  • Deepfake images - AI-generated

Uses proprietary Active Flash 3D technology: emits invisible light flashes and analyzes skin reflection to verify it is real living tissue.

Accuracy: 99.7%
Speed: < 0.3s
Technical Specifications
  • Technology: Active 3D + Passive 3D + Texture Analysis
  • Components: CNN + Depth Estimation + Micro-Texture
  • Training: 5M+ real & fake images
  • Time: < 0.3 seconds
Face Embedding & Digital Fingerprint
Digital Fingerprint

Face Embedding & Digital Fingerprint

Converting facial features into a unique 512-dimensional numeric vector - the Face Fingerprint

Every face processed by ARMA is converted into a unique digital fingerprint (Face Embedding) - a numeric vector of 512 decimal values. Like a digital fingerprint: unique to each person and irreversible to reconstruct the face image.

Comparison between two faces is done by calculating Euclidean Distance or Cosine Similarity between their vectors. The smaller the distance, the stronger the match.

This technology enables searching databases of millions of faces in fractions of a second.

Speed: < 0.05s
Technical Specifications
  • Dimensions: 512-dimension vector
  • Distance: Euclidean L2 & Cosine Similarity
  • Verification Threshold: Customizable (default 0.65)
  • Storage: Only 2KB per face fingerprint
Maximum Privacy

Face Embedding Encryption

Advanced protection layer: even if data leaks, original face embeddings cannot be read

Security and privacy are at the core of ARMA design. We developed an advanced encryption layer for face embeddings ensuring that even in worst-case scenarios - if the database leaks - no one can extract original face images or use leaked fingerprints.

We use Partially Homomorphic Encryption allowing comparison operations on encrypted data without decryption.

Plus Salted Hashing for embeddings, making them irreversible.

Speed: تلقائي
Technical Specifications
  • Encryption: AES-256-GCM + RSA-4096
  • Hashing: SHA-512 with unique Salt per embedding
  • Comparison: Homomorphic Encryption
  • Standards: ISO 27001 | GDPR | NIST
Advanced Database-Backed Search
Ultra-Fast

Advanced Database-Backed Search

Specialized search engine designed specifically for face embeddings ensuring maximum speed & performance

Traditional databases are not suitable for high-dimensional vector search. We built a specialized search engine based on:

  • Vector Indexing (FAISS) - ultra-fast vector indexing
  • Approximate Nearest Neighbor (ANN) - approximate nearest search
  • Clustering + Voronoi Cells - intelligent database partitioning
  • PQ (Product Quantization) - vector compression for memory efficiency

Result: searching 10 million faces takes under 0.05 seconds.

Speed: < 0.05s (10M)
Technical Specifications
  • Engine: FAISS + Custom ANN Index
  • Capacity: 100M+ face embeddings
  • Search Method: IVF-PQ + HNSW
  • Memory: 2KB/face + 8KB index
Integration Ready

Ready-to-Use API

Comprehensive RESTful API with SDKs for 8 programming languages - integrate in minutes

We designed a complete RESTful API enabling developers to integrate ARMA technologies into their apps in minutes. Includes:

  • Face Detection & Landmarks
  • Identity Verification - 1:1 Face Verification
  • Database Search - 1:N Face Search
  • Liveness Detection
  • Age, Gender & Expression Analysis
  • Face Database Management - CRUD operations

We provide SDKs for: Python, JavaScript, Java, C#, Go, PHP, Swift, Kotlin.

Speed: < 100ms
Technical Specifications
  • Protocol: REST + WebSocket for live streaming
  • Format: JSON | Multipart Form Data
  • Auth: API Key + JWT + OAuth2
  • Docs: Swagger/OpenAPI 3.0
  • SDKs: 8 programming languages
Benchmark Results

99.9% Accuracy

Verified Results in the Toughest Conditions

Our model has been tested on global benchmark datasets achieving:

99.9%

Ideal Conditions

98.7%

Low Light

98.2%

Side Angles

97.8%

With Masks

99.9% Accuracy

Want to Experience ArmaNet Accuracy?

Book a dedicated technical demo with our expert team

Request Technical Demo