Skip to content

SCP Federated Training Protocol Specification v0.1(Detailed Version with Contribution Types)

Privacy-preserving Distributed Federated Training Protocol(完整详细规范,含6种贡献类型)

版本:v0.1
状态:Production Protocol Specification(Mainnet-ready Detailed Version)
发布日期:2026-02
适用范围:SCP Compute Plane(Vault / Coordinator / Registry Control Plane)


一、协议目的(Purpose)

SCP Federated Training Protocol(FTP)定义了在 SCP 网络中执行隐私保护分布式模型训练的完整机制。

FTP 的目标是:

  • 支持在 Vault 本地执行模型训练
  • 原始数据永不离开 Vault
  • 支持多种训练贡献类型(gradient、weight_update 等)
  • 支持可验证训练(Proof-of-Training)
  • 支持 extensible schema
  • 支持所有 AI 模型类型

FTP 是 SCP Compute Plane 的核心协议。


二、Training Contribution 抽象模型(Training Contribution Abstraction)

SCP 不限制 Vault 必须返回 gradient。

Vault 返回统一抽象结构:

TrainingContribution

定义:

json
{
  "job_id": "...",
  "vault_id": "...",
  "epoch": 1,
  "contribution_type": "...",
  "contribution": {...},
  "proof": {...}
}

字段说明:

job_id:训练任务 ID

vault_id:Vault ID

epoch:训练 epoch

contribution_type:贡献类型

contribution:贡献 payload

proof:训练证明


三、支持的 6 种 Training Contribution 类型(核心新增章节)

SCP v0.1 明确定义支持以下 6 种 contribution_type:

类型返回内容适用场景是否常用
Gradient梯度 ∇L深度学习、NN、Transformer⭐⭐⭐⭐⭐
Weight Update权重增量 ΔWFedAvg、传统 FL⭐⭐⭐⭐⭐
Full Model Weights完整模型参数小模型、LoRA⭐⭐⭐
Embedding Updateembedding 向量推荐系统、RAG⭐⭐⭐⭐
Sufficient Statistics统计量线性模型、逻辑回归⭐⭐⭐⭐
Custom Contribution自定义RL、强化学习等⭐⭐

3.1 Gradient Contribution

类型:gradient

描述:Vault 返回 loss function 的梯度。

适用模型:Neural Networks、Transformers、Deep Learning models

示例:

json
{
  "contribution_type": "gradient",
  "contribution": [0.0023, -0.0008, 0.0011]
}

Coordinator 聚合:

global_gradient = average(all gradients)


3.2 Weight Update Contribution

类型:weight_update

描述:Vault 返回模型权重更新 ΔW。

适用模型:Federated Averaging(FedAvg)、Neural networks

示例:

json
{
  "contribution_type": "weight_update",
  "contribution": [0.0003, -0.0001, 0.0005]
}

Coordinator 聚合:

global_weights = average(weight_updates)


3.3 Sufficient Statistics Contribution

类型:statistics

描述:Vault 返回 sufficient statistics,而不是 gradient。

适用模型:Linear Regression、Logistic Regression、Gaussian models

示例:

json
{
  "contribution_type": "statistics",
  "contribution": {
    "xtx": [[12.3, 4.5], [4.5, 9.8]],
    "xty": [3.4, 5.6],
    "count": 120
  }
}

Coordinator 可直接计算模型参数。


3.4 Embedding Update Contribution

类型:embedding_update

描述:Vault 返回 embedding 向量更新。

适用模型:Recommendation systems、Embedding models、Vector models

示例:

json
{
  "contribution_type": "embedding_update",
  "contribution": {
    "embedding_delta": [...]
  }
}

3.5 Adapter Update Contribution

类型:adapter_update

描述:Vault 返回 adapter / LoRA 参数更新。

适用模型:LLM fine-tuning、Adapter-based models

示例:

json
{
  "contribution_type": "adapter_update",
  "contribution": {
    "lora_A": [...],
    "lora_B": [...]
  }
}

3.6 Custom Contribution

类型:custom

描述:Vault 返回自定义训练贡献。

适用场景:Reinforcement Learning、Specialized models

示例:

json
{
  "contribution_type": "custom",
  "contribution": {...}
}

四、Contribution Aggregation Rules(聚合规则)

Coordinator 根据 contribution_type 使用不同 aggregation:

gradient → average gradient

weight_update → average weight updates

statistics → aggregate sufficient statistics

embedding_update → merge embeddings

adapter_update → merge adapter weights

custom → use custom aggregation


五、Proof-of-Training(训练证明)

Vault 必须提供 proof:

json
{
  "contribution_hash": "...",
  "vault_signature": "..."
}

Coordinator 验证 proof。


六、安全模型(Security Model)

Vault:never exposes raw data

Coordinator:cannot access Vault data

Registry:stores metadata only


七、隐私模型(Privacy Model)

FTP 保证:

local training only

secure aggregation

proof verification


八、Summary(总结)

SCP Federated Training Protocol v0.1 支持:

6 种 training contribution 类型

privacy-preserving training

verifiable training

scalable federated training

FTP enables decentralized AI training for all model types.