Architecture

SoraChain AI is designed as a privacy-first, decentralized AI training infrastructure. Its architecture integrates Federated Learning with Blockchain, enabling secure, verifiable, and collaborative model development across edge devices. The architecture is built around the Collaborative Model Update (CMU) framework and operates through three primary layers inside interoperable subnets.

1. Global Workflow Layer (Left Section of Diagram)

This layer showcases the complete lifecycle of model training and validation in a decentralized setting.

• Local Training on Edge Devices Devices such as smartphones, wearables, and IoT systems train models locally on private data.

• Global Model on IPFS The Global model is uploaded to IPFS for decentralized, redundant storage. Devices joining in between can directly pull from ipfs to reduce communication impact with the Aggregator

• Blockchain for Metadata Critical metadata (e.g., node identity, timestamp, proof-of-work/proof-of-training) is immutably recorded on the blockchain to ensure auditability and traceability.

• Global Model Update The aggregator uses contributed parameters to update the global model, which is then redistributed to all nodes for the next training round.

---

2. Encrypted Local Device Training Layer (Middle Section)

This section details peer-to-peer synchronization and private training logic.

• Secure Device Alignment Client nodes (e.g., A and B) perform secure handshakes and model alignment using public-private key encryption.

• Data Stays Local No raw data leaves the device. Training occurs on-device using sensitive, high-quality data unavailable to traditional pipelines.

• Encrypted Model Sharing Trained model weights are encrypted and uploaded for inclusion in the global model—preserving full data sovereignty.

---

3. Secure Aggregation Layer (Right Section)

This focuses on how the system safely combines model updates without compromising privacy.

• Public Key Exchange Each node generates and exchanges keys to enable secure gradient exchange.

• Local Computation Each device calculates its own gradients, losses, and model deltas privately.

• Encrypted Aggregation via Aggregator Node Updates are sent to an Aggregator Node, which aggregates encrypted values and updates the global model without ever seeing individual data.

---

System Highlights

• ✅ Zero Raw Data Movement

• ✅ Cryptographically Secure Communication

• ✅ Full Model Auditability on Blockchain

• ✅ Global Coordination, Local Control

This architecture bridges the gap between high-quality private data and collaborative model development while offering unmatched transparency and data integrity.