AI Framework Support

The iExec Platform provides comprehensive support for popular AI and machine learning frameworks. Deploy confidential AI with ease. iExec supports popular AI/ML frameworks, running in secure Trusted Execution Environments (TEEs) with optimized configurations.

Quick Start

Want to get started immediately?

• 📚 AI Frameworks Hello World - Ready-to-use Docker examples for TensorFlow, PyTorch, and more
• 🛠️ Build & Test - General iApp development guide (not AI-specific)
• 🔬 TDX App Guide - Build TDX applications (works well for AI workloads)

Why iExec for AI?

Confidential Computing

Trusted Execution Environments (TEEs) protect your AI models and data end-to-end:

• Data Privacy: TEEs isolate AI computations in secure enclaves
• Secure Training & Inference: Unauthorized entities can never access your models and data
• Hardware-Level Security: Intel SGX and TDX provide enterprise-grade protection

AI Monetization

Monetize your AI assets easily and securely:

• Datasets: Encrypt and sell access to your training data
• Models: Deploy and monetize your trained AI models
• Agents: Create and sell AI agents and applications
• Ownership Preserved: Your digital assets always remain yours

Decentralized Infrastructure

Scale AI applications without centralized cloud dependencies:

• On-Demand Compute: Access powerful resources when you need them
• Fair Pricing: Blockchain verifies execution costs transparently
• Global Network: Deploy across a worldwide network of secure workers

AI Framework Support

Overview

Framework	TDX Support	SGX Support	Best For
TensorFlow	✅ Yes (3.01GB)	❌ No	Deep learning, production ML
PyTorch	✅ Yes (6.44GB)	❌ No	Research, computer vision
Scikit-learn	✅ Yes (1.18GB)	✅ Yes (1.01GB)	Traditional ML, data analysis
OpenVINO	✅ Yes (1.82GB)	❌ No	Computer vision, inference
NumPy	✅ Yes (1.25GB)	✅ Yes (1.08GB)	Scientific computing
Matplotlib	✅ Yes (1.25GB)	✅ Yes (1.08GB)	Data visualization

Framework Details

Framework	Version	Description	TDX Support	SGX Support	Use Cases	Resources
TensorFlow	2.19.0	Google's ML framework for production AI	✅ 3.01GB	❌ Too large	Deep learning, CV, NLP	Docs • Quickstart • Docker
PyTorch	2.7.0+cu126	Facebook's research-focused DL framework	✅ 6.44GB	❌ Too large	Research, DL, CV, NLP	Docs • Quickstart • Docker
Scikit-learn	1.6.1	Comprehensive ML library for Python	✅ 1.18GB	✅ 1.01GB	Classification, regression, clustering	Docs • Examples • Docker
OpenVINO	2024.6.0	Intel's high-performance AI inference toolkit	✅ 1.82GB	❌ Execution issues	Computer vision, inference	Docs • Tutorial • Docker
NumPy	2.0.2	Fundamental package for scientific computing	✅ 1.25GB	✅ 1.08GB	Scientific computing, data analysis	Docs • User Guide • Docker
Matplotlib	3.9.4	Comprehensive library for data visualization	✅ 1.25GB	✅ 1.08GB	Data visualization, plotting	Docs • Gallery • Docker

Getting Started with Docker Examples

What's Included

Our AI Frameworks Hello World repository includes ready-to-use examples:

ai-frameworks-hello-world/
├── tensorflow/     # TensorFlow 2.19.0 example
├── pytorch/        # PyTorch 2.7.0+cu126 example
├── scikit/         # Scikit-learn 1.6.1 example
├── openvino/       # OpenVINO 2024.6.0 example
├── numpy/          # NumPy 2.0.2 example
└── matplotlib/     # Matplotlib 3.9.4 example

Quick Start Commands

# Clone the repository
git clone https://github.com/iExecBlockchainComputing/ai-frameworks-hello-world.git
cd ai-frameworks-hello-world

# Try TensorFlow example
cd tensorflow
docker build -t hello-tensorflow .
docker run --rm hello-tensorflow

# Try PyTorch example
cd ../pytorch
docker build -t hello-pytorch .
docker run --rm hello-pytorch

Features

• ✅ Isolated Testing: Each framework runs in its own container
• ✅ Reproducible: Consistent environment across systems
• ✅ TDX Ready: All containers tested for Intel TDX compatibility
• ✅ Easy Deployment: Simple build and run commands

Technology Comparison

TDX vs SGX for AI

Feature	Intel TDX	Intel SGX
Memory Limit	Multi-GB+	~1.95GB
Framework Support	All major frameworks	Limited (Scikit-learn, NumPy)
Code Changes	Minimal ("lift and shift")	Significant modifications required
Production Ready	✅ Yes	⚠️ Limited
AI Workloads	✅ Excellent	❌ Restricted

Recommendations

For Production AI Applications

• Use TDX for TensorFlow, PyTorch, and OpenVINO
• Use SGX for lightweight ML with Scikit-learn and NumPy

For Development and Testing

• Start with SGX for simple ML tasks
• Migrate to TDX for complex AI workloads

Important Considerations

• SGX Limitations: Expect potential library incompatibilities and code modifications
• TDX Advantages: Minimal code changes required ("lift and shift" approach)

Next Steps

Learn TEE Technologies

• Intel SGX Technology - SGX limitations and capabilities
• Intel TDX Technology - TDX advantages for AI
• SGX vs TDX Comparison - Detailed comparison

Build AI Applications

• Build & Test - Build and test your AI application
• Deploy & Run - Deploy and run your AI application
• Build Intel TDX App - TDX applications for AI workloads
• Inputs - Handle data inputs
• Outputs - Handle data outputs flow in TEE environment

Explore Examples

• AI Frameworks Hello World - Ready-to-use Docker examples
• iExec Discord - Community support