Part 3: Technical Architecture

PlatON does not implement the entire privacy-preserving AI technology stack, focusing on the combination of privacy-preserving computation and AI. The overall architecture is shown in Figure 7, followed by a detailed description of each module.

Privacy-preserving AI Framework (Rosetta)

Rosetta provides privacy-preserving solutions for artificial intelligence where developers do not need expertise in cryptography, federated learning, and trusted execution environments.

• Rosetta integrates mainstream privacy-preserving computation technologies, including cryptography, federated learning, and trusted execution environments, and provides libraries of privacy-preserving statistical analysis algorithms, privacy-preserving machine learning algorithms such as regression, decision trees, and clustering, and privacy-preserving deep learning algorithms such as CNN and RNN.
• Rosetta can be combined with mainstream machine learning and AI frameworks such as TensorFlow, Pytorch, Spark, Flink, etc. Rosetta currently implements the combination with TensorFlow and reuses the TensorFlow API, allowing the migration of legacy TensorFlow code to a privacy-preserving approach with minimal changes.

Underlying Protocol and Privacy-Preserving Computation Protocol on Layer1

Existing public chains do not meet the needs of privacy-preserving AI well, so it is still necessary for PlatON to implement a complete Layer1 base protocol that is deeply adapted to privacy-preserving AI.

P2P Network

PlatON has implemented the P2P base protocol, which is mainly used for node discovery and connection. As a decentralized computing network, there is also a need for discovery and use of data and computing resources, as well as discovery and transparent invocation of AI model services, all of which will be implemented in PlatON 2.0 through the RELOAD protocol.

Giskard consensus algorithm

Giskard is a BFT style consensus optimized in several ways to reduce complexity while further improving throughput through concurrency, with the advantage of high performance and low latency.

• Three-stage Pipeline validation: After the previous block completes a round of voting, it can move on to the next block, and the final confirmation of a block requires the completion of the previous three block votes.
• Concurrent block production and validation: Separate block production and confirmation, concurrently process in Prepare, Pre-Commit and Commit phases.
• Communication optimization: Adopt aggregated signatures to reduce the communication traffic, and also provide an optimized version based on leader to further reduce the communication complexity.
• View-change optimization: Integrate the view-change process into the normal process, eliminating the need for a separate view-change process.

PPoS Economic Model

PPoS is a staking economic model in which every LAT holder can participate. Any node that locks more than a pre-determined minimum number of LATs becomes an alternative node candidate, other LAT holders can lock LATs delegated to alternative node candidates, and the top candidates with the highest number of votes become alternative nodes. After the validators are randomly selected from the alternative nodes using VRF, the validators can participate in block producing and validation. The validators can receive block rewards and transaction fees. The validators and the alternative node share the staking rewards with their supporters according to the prior agreement.

Dual Virtual Machine Support

PlatON supports both EVM and WASM virtual machines and is compatible with solidity contracts. Smart contracts on Ethernet can be ported to PlatON with minor modifications.

Privacy-preserving Smart Contracts

Both the EVM and WASM virtual machines have built-in privacy-preserving algorithms (including homomorphic encryption and zero-knowledge proofs) that developers can use directly in smart contracts to protect the privacy of data within the contract. Based on the privacy-preserving algorithms, PlatON has developed a standard for privacy token contracts that incorporates minting, destruction, and interaction with standard tokens to anonymize them.

Privacy-preserving computation network (Metis)

Metis aggregates the data, algorithms, and computing power needed for computing in a decentralized manner to create a secure privacy-preserving computation paradigm.

Decentralized Scheduling

Layer2 underlying network is RELOAD overlay network, data nodes and computing nodes are connected through P2P protocol, and the RELOAD protocol is used to publish, discover, locate and schedule data and computing resources.

Data Service

The data subject can either start the data node locally or host the data encrypted to the data node. Upon receiving a computation request, the data node uses secret sharing to slice the data and distribute it to randomly selected computing nodes for secure multi-party computation. The computation task and selection of computing nodes need to be confirmed among multiple data nodes by consensus protocol. Data nodes can also encrypt the data by homomorphic encryption, distribute it to computing nodes for outsourced computation, and verify the returned computation results and computation proofs using verifiable computation algorithms.

Computation Service

Metis supports two different types of privacy-preserving computation protocols and can also be extended with additional privacy-preserving computation protocols.

secure multi-party computation

The privacy-preserving computation is performed between computation nodes following the secure multi-party computation protocol, and the computation results are returned to the computation result party through blockchain smart contracts. In the case of AI model training, the completed AI model can be deployed to Layer3’s AI network and become an AI agent to provide AI services to the outside world.

secure outsourcing computation

If users have their own data and algorithms, but do not have enough computing power, they can give their data (after homomorphic encryption) and algorithms to third-party computing nodes for outsourced computation. The data and algorithms can be distributed to multiple compute nodes for parallel computation, and the computation task can be decomposed according to the data or model. After the computation is completed, the computing nodes return computation results and computation proofs, which can verify the correctness of the computation.

Blockchain

Establish decentralized data, arithmetic and model trading market on blockchain network through economic incentives and smart contracts of blockchain. The economic model of privacy-preserving computation is implemented to assetize and monetize data and arithmetic resources. To ensure the security and validity of data and computation, the economic model contains staking and slash mechanisms. All data, variables and processes used in privacy-preserving computation have tamper- evident records that can be tracked and audited.

Decentralized identity (DID) schemes are used to enable decentralized authentication and authorization of nodes and resources, including data validation and usage authorization. DID refers to a set of fully decentralized, allowing us as individuals or our own organizations to have full ownership, management and control over their identity.

Privacy-preserving AI Platform (Moirae)

Moirae is a decentralized AI “cloud” platform that provides a one-stop AI development platform for developers, offering a one-stop, all-inclusive modeling process to help users quickly create and deploy models and manage full-cycle AI workflows. On the other hand, it is also an open AI marketplace where developers can access training datasets on the network to train models, launch AI models, and interact with other AI models and paid users.

All-in-one AI development platform

• Full range of modeling processes

Moirae integrates data import, data processing, model development, model training, model evaluation, and service go-live to provide a one-stop, all-inclusive machine learning and deep learning modeling process to quickly build intelligent businesses. It provides visual and low-code development tools, automated model generation, and continuous training and deployment for machine learning and deep learning, lowering the threshold of developers, helping users to quickly create and deploy models and manage full-cycle AI workflow.

• Distributed model training and service hosting

Metis provides globally distributed computing power and supports multiple chip architectures such as GPU and FPGA at the AI computing level, forming a heterogeneous AI computing platform that allows AI developers to directly submit computational tasks such as data pre-processing, feature engineering, and model training at low cost, with computational resources automatically scheduled on demand.

Moirae also provides service hosting. Trained AI models can be deployed directly to a single network node, or can be deployed in pieces to multiple network nodes, with multiple nodes making predictions through secure multi-party computing protocols.

Open Artificial Intelligence Marketplace

• Data marketplace

Moirae builds a data exchange protocol based on zero-knowledge proofs and fair exchange protocol through which training datasets can be traded fairly and no party can gain an advantage by early withdrawal or other misbehavior. Training datasets are not exchanged explicitly, but rather participate in model training through secure multi-party computation protocols.

AI developers can actively search for training datasets in the data marketplace, or they can publish models and let others provide data to collaborate on training models.To protect the privacy and security of the data, the data needs to be authorized and trained with models through privacy- preserving outsourced computation or secure multi-party computation protocols.

Data marketplaces create incentives through cryptomics to encourage the submission of data to improve the accuracy of models. To ensure the validity of data, data providers are required to stake and penalize when verified as bad data.

• AI Service Marketplace

The trained models can be deployed directly on the network and provide prediction services to the public. The prediction service information is registered in a smart contract, which can be searched and invoked by paying users.

Autonomous AI Agent Network (Horae)

For developers who want to provide AI services over the web, the most critical component is the service node. A service node is an execution container for an AI model that can host multiple AI models and provide AI services to the outside world. Considering network redundancy and fault tolerance, AI models can generally be hosted to multiple service nodes and can be migrated between service nodes.

Registry nodes and evaluation nodes form an intelligent search network that facilitates service and agent matching and exchange. AI services and agents register their text descriptions and labels to registry nodes so that users can discover their services, pricing, addresses, and other information and invoke them. Evaluation nodes perform service testing, evaluation, and rating of AI services and agents, and collectively maintain a reputation scoring system through consensus algorithms, which is used as a basis for search and recommendation, enabling other users to query AI services and agents quickly and easily. The effectiveness of evaluation can be improved by machine learning algorithms trained on historical data, such as successful queries and interactions. Using machine learning-based search, users are able to identify potential AI services and agents.

Horae aims to use self-organizing group intelligence to create a whole that is greater than the sum of its parts. Autonomous agents do not just exist in the digital world, but can also serve as a bridge between the digital world and the real world, connecting to humans, IoT devices, and external IT systems. Each autonomous agent is a daemon that operates independently, each pursuing its own relatively simple goals, but in their interaction will generate more complex goals and generate more intelligent higher-order agents.

For AI agents to be truly autonomous, they need an understanding of how to talk to each other that goes far beyond merely knowing the appropriate communication protocols. The application of Natural Language Processing (NLP) and Process Mining technologies enables autonomous agents to understand tasks described in natural language and develop true autonomy.