Process

Process Overview: (1a) Task Creation; (1b) Registration; (2) Initialization; (3)-(5) Local Training; (6)-(8) Evaluation & Aggregation; (9) Rewards Distribution; (10) Model Inference

1. Task Creation and Registration

   1. Task Creation: Task Publisher creates a new Task and publishes it to the blockchain. Model’s parameters (and possibly additional task attributes) and the Genesis Model are stored on IPFS. The returned content identifier (CID) is recorded on the blockchain.

   2. Registration: Eligible Trainers and Aggregators review the training task via a dedicated decentralized web application (including device specifications, task requirements, potential benefits, etc.) and register for the task. Registration can also be subscription-like, where they register once and automatically participate in all relevant tasks. Depending on the task, the device maintains an on-device repository of data relevant to the training process or accesses an external data repository (public or privately accessible to Trainers, given a key from the Task Publisher). During registration, the client installs the associated Training Agent (if necessary).

2. Initialization: When all conditions for the task are met, the Task Manager Smart Contract initiates the task, and it is broadcasted to registered Trainers and Aggregators.

3. Local Training [Download]: Each device downloads the latest version (at time t) of the CID of the Global Model from the blockchain and the model parameters from IPFS.

4. Local Training: The Training Agent deployed on (possibly randomly selected) Trainers’ devices, invokes a background training process. Local training continues until the end condition is met.

Local Training

1. Local Training [Upload]: Following the local training, the local model (Proposed Weights Vector) undergoes Differential Privacy and Encryption processes, then is uploaded to IPFS. It Content Identifier (CID) is sent via a transaction and stored on the blockchain.

2. Evaluation & Aggregation [Download]: A random aggregator is selected for each aggregation round, triggered by the creation of new blocks, based on the Committee-Based Consensus Algorithm. The frequency of elections can be adjusted based on specific needs.

3. Evaluation & Aggregation: Selected Aggregators (which may also include a subset of the Trainers in supported scenarios) download a predefined dataset from IPFS and evaluate the trainers, scoring Trainers for accuracy. The Aggregator interacts with the smart contract to retrieve lists of the trainers, model CIDs, and contribution points for the recorded models.

4. Evaluation & Aggregation [Upload]: The new global model is uploaded to IPFS, while its CID and the evaluations of Trainers are sent to the blockchain.

5. Rewards Distribution: Rewards are distributed to Trainers and Aggregators based on their evaluated contribution.

[Repeat steps 3 – 9 until end conditions are met]

1. Model Inference: Once the training process converges, the Task Publisher downloads the latest global model parameters from IPFS and utilizes them for their needs.