Nillion AIVM

::: warning Nillion AIVM is currently in early development. While functional, it may contain bugs and is not recommended for deployment in production or critical systems. ::: Nillion AIVM is a secure inference platform for Deep Learning models based on Multi-Party Computation (MPC). It enables private model inference and custom model deployment while maintaining data confidentiality throughout the computation process. This documentation covers installation, supported models, and usage instructions.

Supported Models

AIVM currently supports the following pre-trained models for specific learning tasks. You can either utilize these existing models or deploy your own custom-trained versions:

BertTiny

• SMS Spam Classification
  • Binary classification for detecting spam messages
  • Input: Text string
  • Output: Binary classification (spam/not spam)
• Movie Sentiment Analysis
  • Sentiment analysis for movie reviews
  • Input: Text string
  • Output: Sentiment score (-1 to 1)

LeNet5

• Handwritten Digit Recognition (MNIST)
  • Classification of handwritten digits
  • Input: 28x28 grayscale image
  • Output: Digit classification (0-9)
• Cats vs Dogs Classification
  • Binary image classification
  • Input: 28x28 grayscale image
  • Output: Binary classification (cat/dog)

Installation

Installing Nillion AIVM is straightforward:

pip install nillion-aivm

This command installs all necessary dependencies for performing secure inference on AIVM.

Starting the Development Network

Launch the AIVM development network with:

aivm-devnet

This command starts a persistent process that manages the secure computation infrastructure. To stop the network, use CTRL+C.

::: info

Note: Ensure aivm-devnet is running before proceeding with the following examples.

:::

Performing Secure Inference

Basic Usage

1. First, import the AIVM client and check available models:

import aivm_client as aic

# List all supported models
available_models = aic.get_supported_models()
print(available_models)

2. Prepare your input data. Here's an example using PyTorch to generate a random input:

import torch

# Create a sample input (e.g., for LeNet5 MNIST)
random_input = torch.randn((1, 1, 28, 28))  # Batch size 1, 1 channel, 28x28 pixels

3. Encrypt your input using the appropriate Cryptensor:

# Encrypt the input
encrypted_input = aic.LeNet5Cryptensor(random_input)

4. Perform secure inference:

# Get prediction while maintaining privacy
result = aic.get_prediction(encrypted_input, "LeNet5MNIST")

The get_prediction function automatically handles the secure computation protocol with the aivm-devnet nodes, ensuring that your input data remains private throughout the inference process.

Custom Model Deployment

You can deploy your own trained models to AIVM, provided they follow the supported architectures (BertTiny or LeNet5).

Uploading Custom Models

1. Import the AIVM client:

import aivm_client as aic

2. Upload your custom model:

# For BertTiny models
aic.upload_bert_tiny_model(model_path, "MyCustomBertTiny")

# For LeNet5 models
aic.upload_lenet5_model(model_path, "MyCustomLeNet5")

3. Perform inference with your custom model:

# For BertTiny models
result = aic.get_prediction(private_berttiny_input, "MyCustomBertTiny")

# For LeNet5 models
result = aic.get_prediction(private_lenet5_input, "MyCustomLeNet5")

Model Requirements

Custom models must meet these requirements:

• Follow the exact architecture of either BertTiny or LeNet5
• Be trained using PyTorch
• Use the same input dimensions as the original architectures
• Be saved in PyTorch's standard model format (.pth or .pt)

info

Note: Custom model names must be unique within your AIVM instance.