CRT Federated Learning

Google/Apple FL patents. CC0 alternative.

Federated learning trains models across distributed participants without sharing raw data. Standard FL aggregates full model updates (expensive, gradient leakage risk). CRT FL: each of 7 participants trains one channel. Aggregation = CRT reconstruction. Four parity channels (mod 9, 11, 13, 17) give free poison detection. Block-diagonal: zero cross-participant gradient leakage. Privacy by algebra, not by protocol.

How It Works

CRT Federated Theorem
7 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, 13, or 17). The server collects 7 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. Three data channels (mod 8, 25, 49) carry model parameters; four parity channels (mod 9, 11, 13, 17) verify integrity.
7 participants
7 CRT channels
Each trains independently on private data. Shares only a tiny weight.
Block-diagonal
Zero gradient leak
Poisoning one channel cannot affect another. Algebraic firewall.
Quad parity
Free poison check
mod-9, mod-11, mod-13, mod-17 channels detect corrupted participants.
Data + parity
3 + 4 split
3 data channels carry the model. 4 parity channels verify integrity.

Try It

Seed (sets true weight):

7 participants each learn their channel's weight from private data. CRT reconstructs the global model. All 7 match = exact reconstruction.

Poison Detection

Injects 80% corrupted data into the mod-11 participant. Compare clean vs poisoned weights per channel.

CRT FL vs Standard Federated Learning

AggregationFedAvg: weighted average of full model updatesCRT: reconstruct from 7 tiny channel weightsCommunicationFull gradient vectors (millions of params)7 integers (max value 48). Minimal.PrivacyGradient leakage attacks possibleBlock-diagonal: zero cross-channel gradient accessPoison detectionRequires separate anomaly detection systemmod-11 deviation = instant free detectionFault toleranceOne bad participant corrupts whole modelOne bad channel = 1/7 of model. Others unaffected.Patent statusGoogle (on-device), Apple (differential privacy FL)CC0. Public domain. Forever.

Source code · Public domain (CC0)

Report issue

.ax source compiled to WASM via self-hosting compiler. Zero HTML authored.