mirror of https://github.com/zoo-apps/ZIPs.git synced 2026-10-30 14:57:22 +00:00

Files

T

zeekayandHanzo Dev d1410729da Publish under the org, not a person

Bylines, paper authorship, correspondence addresses and the founder blurb under
each post named an individual. They name the organisation that published the
work instead: the brand on posts and pages, the research byline on papers, and
research@ in place of a personal address. Both are what these repositories
already used elsewhere — nothing new was coined.

Co-authors are untouched. Where a paper named someone else alongside, they stay
named; only the one byline changed. Bibliography entries, running heads,
PDF metadata and affiliations move with it, so a citation still resolves to
whoever the cited paper names.

Sample and fixture data uses the placeholder name these tests already use
elsewhere, so no assertion changed meaning.

Co-authored-by: Hanzo Dev <[email protected]>

2026-09-01 02:31:34 -07:00

5.6 KiB

Raw Permalink Blame History

zip, title, description, author, authors, status, type, category, originated, traces-from, follow-on, created, tags, requires, references, license

zip

title

description

author

authors

status

type

ZIP-0410: Decentralized Semantic Optimization (DSO)

Abstract

This proposal specifies the Decentralized Semantic Optimization (DSO) protocol, the distributed counterpart to ASO (ZIP-0409). DSO enables geographically dispersed nodes to collaboratively improve shared AI models by exchanging semantic gradients -- compressed, privacy-protected representations of local learning signals -- rather than raw data or full parameter updates. The protocol guarantees differential privacy at configurable (epsilon, delta) budgets, employs Byzantine-robust aggregation, and records all training contributions on an immutable ledger for provenance and royalty distribution. DSO is the foundational training protocol for all Zoo AI systems.

Motivation

ASO (ZIP-0409) optimizes what a model learns; DSO optimizes how multiple parties collaborate on learning without compromising data privacy. This is critical because:

Privacy: Conservation organizations cannot share sensitive wildlife data (endangered species locations, poaching coordinates) without risking exposure
Data sovereignty: Data collected on indigenous lands or within national parks is subject to legal restrictions prohibiting export
Compute democratization: DSO enables small conservation groups to contribute to model training using their own hardware (ZIP-0407)
Attribution: Every training contribution is recorded on-chain, enabling fair credit and royalty distribution

Specification

System Architecture

+------------------+     +------------------+     +------------------+
|   DSO Node A     |     |   DSO Node B     |     |   DSO Node C     |
| (Camera Traps)   |     | (Acoustic Data)  |     | (Satellite Imgs) |
|                  |     |                  |     |                  |
| Local Data Store |     | Local Data Store |     | Local Data Store |
| Local Trainer    |     | Local Trainer    |     | Local Trainer    |
| Gradient Encoder |     | Gradient Encoder |     | Gradient Encoder |
+--------+---------+     +--------+---------+     +--------+---------+
         |                         |                         |
         | Semantic Gradients      | Semantic Gradients      |
         | (encrypted, compressed) | (encrypted, compressed) |
         v                         v                         v
+------------------------------------------------------------------+
|                   DSO Aggregation Layer                            |
|  Byzantine-robust aggregation + differential privacy enforcement  |
+------------------------------------------------------------------+
         |
         v
+------------------+
| Updated Model    |
| (IPFS checkpoint)|
| (on-chain CID)   |
+------------------+

Semantic Gradient Protocol

Local training: Each node trains on its local data for K steps
Gradient encoding: Full gradients are compressed into semantic gradients via:
- Low-rank decomposition (rank r << model dimension)
- Quantization to 4-bit representation
- Differential privacy noise injection (calibrated to target epsilon)
Transmission: Encoded gradients are sent to the aggregation layer
Aggregation: Geometric median aggregation (Byzantine-robust)
Model update: Aggregated gradient applied to global model

Privacy Guarantees

Each round satisfies (epsilon, delta)-differential privacy
Privacy budget is tracked cumulatively across rounds
Nodes can set their own privacy level (stricter = more noise = less contribution weight)
Composition theorem bounds total privacy loss over T rounds

Contribution Tracking

Every gradient submission is recorded on-chain:

ContributionRecord {
  node_id: DID
  round: uint64
  gradient_cid: CID       // IPFS hash of semantic gradient
  data_summary: Hash      // commitment to local data statistics
  compute_proof: Proof    // verifiable compute attestation
  privacy_budget_used: float
  timestamp: uint64
}

Research Papers

experience-ledger-dso -- DSO protocol with Experience Ledger integration (2025)
hanzo-dso -- Hanzo DSO implementation specification
zen-dso-protocol -- DSO protocol for Zen model training
zen-distributed-training -- Distributed training infrastructure

Implementation

hanzo/node: Blockchain/AI node with DSO protocol support
hanzo/candle: Rust ML framework for gradient encoding/decoding
zoo/contracts: On-chain contribution tracking contracts

Timeline

Originated: September 2023 (DSO protocol design, combining ZIP-0407 and ZIP-0409)
Research: hanzo-dso published 2023, zen-dso-protocol published 2024, experience-ledger-dso published 2025
Implementation: DSO protocol in Hanzo Node 2024

5.6 KiB Raw Permalink Blame History