OPEN SOURCE · v2.4.1 · 6 NODES ACTIVE

Train everywhere.
Share nothing.

A federated learning framework for ML teams who need model convergence across siloed data — without a single row of raw data leaving its origin node.

cross-silocross-devicehybrid

Start Federating Free →View on GitHub

2.3kGitHub stars

Apache 2.0License

PyTorch · TF · JAXFrameworks

Hospital ABerlin

r47

Hospital BToronto

r45

Bank Node 3São Paulo

r43

Edge ClusterSeoul

r46

Research LabGeneva

r44

Bank Node 1Singapore

r42

ROUND 47 / 64 · AGGREGATING73%

6 nodes · ε=0.82 · δ=1e-5FedAvg+DP

LIVE · CROSS-SILO

6.2×Fewer comm. rounds vs. FedAvg baseline

ε = 0.82Privacy budget per training run

99.4%Accuracy vs. centralized training

<40msAggregation latency per round

CAPABILITY MATRIX

What it does.
How it protects.

Four primitives. Every deployment mode. No raw data in transit — ever.

Differential Privacy

Mathematical guarantees, not promises.

Per-sample gradient clipping with calibrated Gaussian noise injection. Privacy budget tracked across every round — ε and δ visible in real time. Moments accountant for tight composition.

Rényi DP accountantPer-round budget allocationGradient clipping: L₂ normNoise multiplier σ configurable

federate_config.pydiff_privacy

1from federate import DPTrainer, PrivacyAccountant

3trainer = DPTrainer(

4 model=my_model,

5 noise_multiplier=1.1,

6 max_grad_norm=1.0,

7 target_epsilon=1.0,

8 target_delta=1e-5,

11accountant = PrivacyAccountant(trainer)

12# ε=0.82 after round 47 of 64

13print(accountant.spent_budget()) # → {ε: 0.82, δ: 1e-5}

Secure Aggregation

The server learns the sum. Nothing else.

aggregator.py

Asynchronous Rounds

Slow nodes don't block fast ones.

round_config.py

Model Heterogeneity

Different architectures, one global signal.

hetero_fed.py

PRIVACY GUARANTEES

How it protects.

Privacy is not a checkbox. Every Federate training run produces a signed privacy receipt — ε spent, δ bound, rounds completed, nodes that contributed. Auditable. Reproducible. Court-admissible.

∂

Differential Privacy

ε = 0.82per training run

Rényi divergence accountant tracks cumulative privacy loss across all 64 rounds. Budget never exceeds configured ε=1.0 threshold.

∑

Secure Aggregation

0 bytesraw gradients exposed

Shamir (4,6)-threshold secret sharing. Aggregation server reconstructs only the weighted average — individual updates remain cryptographically masked.

⊕

Communication Compression

6.2×fewer rounds

Top-k gradient sparsification with error feedback. Sends only the most informative 5% of parameters per round without convergence penalty.

⊗

Homomorphic Layer

100%model update encryption

Optional CKKS-based homomorphic encryption layer for regulatory environments requiring encrypted aggregation (HIPAA, GDPR Art. 25).

PRIVACY RECEIPT · RUN #4719 · 2026-02-25T10:10:37Z

ε spent0.82

δ bound1e-5

rounds47/64

nodes6

strategyFedAvg+DP

statusACTIVE

CRYPTOGRAPHICALLY SIGNED

DEPLOYMENT MODES

Where it deploys.

Three topologies. One API. Deploy to institutional silos, edge device fleets, or mixed hierarchical networks without changing your training code.

Berlin

Toronto

São Paulo

Geneva

Cross-Silo

Healthcare · Finance · Research

Dozens of powerful institutional nodes — hospitals, banks, universities. High-trust, low-frequency rounds. Full differential privacy and secure aggregation on every update.

6–200 institutions

TLS mutual auth

SOC2 / HIPAA ready

Round: minutes

99.4%accuracy vs. centralized

Seoul

Lagos

Mumbai

Chicago

Cross-Device

IoT · Mobile · Edge

Millions of resource-constrained edge devices. Asynchronous rounds with staleness tolerance. Partial participation, bandwidth-aware client selection, on-device quantization.

1M+ devices

Async rounds τ≤8

Top-k compression

Round: seconds

6.2×fewer rounds than FedAvg

Berlin

Seoul

São Paulo

Lagos

Hybrid

Mixed Topology

Institutional silos coordinating alongside edge device clusters. Hierarchical aggregation — local aggregators in each silo, global aggregation across silos. Best of both worlds.

Hierarchical agg

Mixed precision

Model heterogeneity

Custom topology

<40msaggregation latency

FRAMEWORK COMPARISON

Federate vs. the field.

Benchmarked on a 6-node cross-silo setup. CIFAR-10 proxy task, IID distribution, 64 rounds max.

CAPABILITY	Federate	PySyft	Flower	NVIDIA FLARE
Differential Privacy (built-in)	✓	✓	—	—
Secure Aggregation	✓	✓	—	✓
Async rounds	✓	—	✓	—
Model heterogeneity	✓	—	—	—
Cross-device support	✓	—	✓	—
Cross-silo support	✓	✓	✓	✓
Privacy receipt / audit log	✓	—	—	✓
Homomorphic encryption	✓	✓	—	—
Framework agnostic	✓	—	✓	✓
Managed cloud option	✓	—	—	—
Convergence (rounds to 95% acc.)	47	112	89	74
Setup time	< 5 min	~2 hours	~30 min	~3 hours

QUICKSTART · NO CREDIT CARD · GITHUB NOTEBOOK

First federated run in 5 minutes.

Open a hosted Jupyter notebook. Connect two sample nodes. Watch the topology animate. No infra required.

Start Federating Free →

Train everywhere.Share nothing.

What it does.How it protects.

Differential Privacy

Secure Aggregation

Asynchronous Rounds

Model Heterogeneity

How it protects.

Differential Privacy

Secure Aggregation

Communication Compression

Homomorphic Layer

Where it deploys.

Cross-Silo

Cross-Device

Hybrid

Federate vs. the field.

First federated run in 5 minutes.

Train everywhere.
Share nothing.

What it does.
How it protects.