The model that gets smarter
every month. Permanently ahead.
Every RMADOR deployment contributes to a shared model that no competitor can replicate. Each site trains locally on its own footage. Only gradient updates — never video — are shared. Flower FedAvg aggregates improvements from the entire fleet into a global model pushed to every deployment monthly.
The compounding math
Federated learning does not improve linearly. Each additional site contributes unique threat scenarios, environmental conditions, and camera angles. The marginal value of the 500th site is greater than the first 100 combined.
2×
better at 100 sites
than training on a single site's footage alone
5×
better at 500 sites
edge cases from 500 distinct environments — impossible to replicate
∞
at 1,000+ sites
structural moat — competitors cannot buy this training signal
Improvement estimates based on FedAvg convergence literature. Actual improvement varies by site diversity and local dataset quality.
How it works
Local training
Each site runs a fine-tuning pass on its own detection events and labelled alerts. Training happens on the local GPU — footage never leaves the premises.
Gradient extraction
Only the model gradient update is extracted — not weights, not data. Opacus applies differential privacy noise (ε = 4.0) before the gradient leaves the device.
FedAvg aggregation
Flower aggregates signed gradient updates from all participating sites using the FedAvg algorithm. Each gradient is Ed25519-signed by the submitting node.
Global model pushed
The improved global model passes production gates (mAP@50-95 ≥ 0.80, bias audit, 48h shadow deployment) then is pushed to all sites via the OTA MQTT channel.
Privacy by design
Federated learning was designed from first principles to make video sharing impossible — not as an afterthought. The architecture guarantees that sensitive footage cannot be reconstructed from the gradient updates even if the aggregation server is compromised.
Elastic Weight Consolidation (EWC) prevents catastrophic forgetting — the global model retains its original 7-class detection accuracy while incorporating the new knowledge from the fleet.
Technical specifications
The compounding competitive advantage
Competitors can copy features. They cannot copy years of real-world training signal from a deployed fleet.
Year 1
Advantage
Model trained on 3–5× more real-world security scenarios than any self-contained system. False alarm rate drops as the model adapts to common false-positive patterns from the fleet.
Year 2
Moat
Competitor models stagnate on their training data. RMADOR's model has seen edge cases from hundreds of sites — attack angles, disguise patterns, coordinated probe behaviours — that no competitor dataset can match.
Year 3
Unassailable lead
A new entrant would need to deploy at scale and wait years to accumulate equivalent training signal. The model advantage is compounding and structural — not replicable by buying more compute.
Opt in from your admin dashboard — zero configuration required.
Federated learning participation is a single toggle in your RMADOR account settings. No infrastructure changes, no API keys, no engineering work. Consent can be revoked at any time — your locally-trained weights remain yours regardless.
Federated learning requires tenant consent per GDPR Art. 4(5). Gradient updates are pseudonymised — no site identity is transmitted with the gradient. You can opt out at any time without affecting your local model version.