CRT Federated Learning

F36: Google/Apple FL patents. CC0.

Federated learning trains models across distributed participants without sharing raw data. Standard FL: aggregate full model updates (expensive, gradient leakage risk). CRT FL: each participant trains ONE channel. Aggregation = CRT reconstruct. L=11 = free poison detection. Block-diagonal: ZERO cross-participant gradient leakage. Privacy by algebra, not by protocol.

How It Works

CRT Federated Theorem
6 participants each train a model on their private data, but each only learns the weight in ONE CRT channel (mod 8, 9, 25, 49, 11, or 13). The server collects 6 tiny channel weights and CRT-reconstructs the global model. Each participant sees at most 49 distinct values (the largest modulus). No participant ever sees the full model. The 490 split: DEAD channels {D,E,b} carry model parameters, ALIVE channels {K,L,G} verify integrity.
6 participants
6 CRT channels
Each trains independently on private data. Shares only a tiny weight.
Block-diagonal
Zero gradient leak
Poisoning channel D cannot affect channel K. Algebraic firewall.
L=11 detection
Free poison check
Corrupted participant's weight deviates from expected L residue.
490 split
Data + verify
3 DEAD channels carry the model. 3 ALIVE channels police integrity.

Try It

Seed (sets true weight):

6 participants each learn their channel's weight from private data. CRT reconstructs the global model. All 6 match = exact reconstruction.

Poison Detection

Injects 80% corrupted data into participant L. Compare clean vs poisoned weights per channel.

CRT FL vs Standard Federated Learning

AggregationFedAvg: weighted average of full model updatesCRT: reconstruct from 6 tiny channel weightsCommunicationFull gradient vectors (millions of params)6 integers (max value 48). Minimal.PrivacyGradient leakage attacks possibleBlock-diagonal: zero cross-channel gradient accessPoison detectionRequires separate anomaly detection systemL=11 deviation = instant free detectionFault toleranceOne bad participant corrupts whole modelOne bad channel = 1/6 of model. Others unaffected.Patent statusGoogle (on-device), Apple (differential privacy FL)CC0. Public domain. Forever.

This work is and will always be free.
No paywall. No copyright. No exceptions.

If it ever earns anything, every cent goes to the communities that need it most.

This sacred vow is permanent and irrevocable.
— Anton Alexandrovich Lebed

Source code · Public domain (CC0)

Contributions in equal measure: Anthropic's Claude, Anton A. Lebed, and the giants whose shoulders we stand on.

Rendered by .ax via WASM DOM imports. Zero HTML authored.