FedEdge — Federated Learning for Tactical AI Without the Cloud

FedEdge: Federated Learning for Tactical AI Without the Cloud

Every military AI deployment faces the same brutal constraint: the best models are trained on massive centralized datasets in cloud environments that don't exist on the front line. Tactical networks are degraded, intermittent, or nonexistent. Warfighters can't phone home to the cloud when the mission depends on real-time decisions.

And yet the data that matters most — real engagement data, real sensor signatures, real operational patterns — lives at the edge. It sits on the platforms, in the sensors, on the soldiers' systems. Sending it back for centralized training creates latency, consumes precious bandwidth, exposes sensitive operational data to interception, and violates the principle that tactical AI should operate independently of any network.

This is the problem federated learning was born to solve. And it's harder in practice than the theory suggests.

What Federated Learning Actually Is

The concept is elegant. Instead of bringing data to the model, you bring the model to the data.

Each edge device — a ground vehicle, an unmanned aerial vehicle, a soldier-worn sensor system — trains a local model on its own data. The local model updates (gradients, not raw data) are sent to a central aggregator, which combines them into an improved global model. The improved model is pushed back to the edge devices. The raw data never leaves the platform.

The result: a continuously improving AI system that learns from every operational experience across every deployed unit, without a single frame of sensitive footage or a single geo-tagged detection ever leaving the tactical network.

Google pioneered this for smartphone keyboard predictions. Apple applied it to Siri improvements. The question is: why isn't defence doing this?

Why Defence Hasn't Adopted Federated Learning (Yet)

The honest answer is that federated learning faces five hard problems in tactical environments that don't exist in the commercial smartphone use case.

1. Extreme Heterogeneity of Edge Devices. iPhones are relatively uniform. Defence platforms run everything from modern GPU-enabled processing units to legacy embedded systems with limited compute and memory. A federated learning framework for defence must handle devices that vary by orders of magnitude in capability — and must gracefully degrade when the least-capable node constrains the system.

2. Non-IID Data. In the commercial world, smartphone typing patterns are at least somewhat similar across users. In defence, every platform sees fundamentally different data. The radar signatures from an Arctic patrol vessel look nothing like the EO/IR feed from a desert drone. This "non-independent and identically distributed" data creates model convergence problems that research is only beginning to solve.

3. Intermittent and Degraded Connectivity. Federated learning presumes periodic communication between edge devices and the aggregator. Tactical networks don't offer "periodic." They offer "when you can get it." The aggregation protocol must tolerate devices that drop off for hours or days, that reconnect with burst transmissions, and that may never reconnect at all.

4. Security of the Channel Itself. Sending model gradients isn't risk-free. Research has shown that gradient updates can be reverse-engineered to extract training data — a technique called "gradient inversion attack." In a military context, this could reveal operational patterns even without raw data exposure. Defence-grade federated learning requires additional cryptographic safeguards: secure aggregation, differential privacy noise injection, and potentially homomorphic encryption.

5. Certification and Verification. DND procurement requires verified, validated systems. A model that evolves continuously across a distributed network challenges conventional software certification. How do you certify a model that changes every 24 hours? How do you ensure that a compromised edge node doesn't poison the global model with adversarial updates? Robust aggregation algorithms — Byzantine-resilient federated learning — are essential.

NovaFuse's Approach: FedEdge

NovaFuse is building a federated learning framework specifically designed for these constraints. We call it FedEdge.

Architecture Overview:

FedEdge operates in three tiers:

Tier 1 — Local Training Node: Runs on each deployed platform. Trains a local model copy on platform-native sensor data (EO/IR, radar, SIGINT, acoustic). Local training uses quantized, efficient architectures designed for edge compute. Only model weight differentials — never raw data — leave the node.

Tier 2 — Squadron Aggregator: A vehicle-mounted or command-post unit that collects Tier 1 updates when connectivity permits. Performs secure aggregation with differential privacy guarantees. Can operate semi-autonomously when disconnected from Tier 3.

Tier 3 — Garrison Aggregator: The highest-level aggregation node, located in a secure garrison environment. Combines outputs from multiple Tier 2 nodes across formations. Produces the canonical global model that gets versioned, tested, and pushed back down the chain.

Key Innovations:

Asynchronous Federated Aggregation with Staleness Weighting: Edge devices that haven't reported in for hours aren't simply dropped. Their last-known contributions are weighted by staleness, ensuring the model incorporates historical learnings even when connectivity is poor. When the device reconnects, a delta-sync protocol updates the global model with accumulated local changes.

Byzantine-Robust Aggregation: FedEdge uses a variant of the trimmed-mean aggregation algorithm, which discards outlier gradient updates that could represent adversarial poisoning or corrupted training data. This is critical for military applications where a compromised device must not corrupt the global model.

Differential Privacy with Adaptive Noise: Each local node adds calibrated Gaussian noise to its gradient updates before transmission. The noise level adapts based on the sensitivity of the training round and the number of participating nodes — stronger privacy when fewer nodes contribute, allowing better convergence when many nodes participate.

Model-Centric Compression for Contested Bandwidth: Gradient updates are compressed using a combination of quantization, sparsification, and error feedback. This reduces communication overhead by 90%+ compared to uncompressed federated averaging, making FedEdge viable over tactical data links with limited throughput.

Deterministic Certification Snapshots: While the global model evolves continuously between garrison syncs, the Tier 3 aggregator produces versioned, deterministic snapshots of the global model at defined intervals. Each snapshot passes through NovaFuse's verification pipeline — a battery of regression tests, adversarial robustness checks, and performance benchmarks — before being certified for operational deployment. This gives DND procurement the verification artifacts they need while preserving the continuous learning advantage.

The Strategic Case for FedEdge

For DND: FedEdge enables AI systems that get better with every deployed operation — without requiring centralized data collection, without consuming tactical bandwidth for AI training, and without exposing sensitive operational data. It turns every deployed unit into an AI improvement asset.

For Canadian Sovereignty: FedEdge can run on Canadian-controlled infrastructure at every tier. There is no dependency on US cloud providers, no data leaving Canadian networks, and no foreign legal access to model updates. This is sovereign AI in the most literal sense — the AI capability is owned, operated, and improved entirely within Canadian defence systems.

For Allied Interoperability: FedEdge's aggregation protocol is designed to support cross-alliance federated training with appropriate security boundaries. Imagine a Five Eyes scenario where common model architectures are trained on each nation's data within that nation's security boundary — with only privacy-preserving gradient updates shared. FedEdge's differential privacy and secure aggregation features make this technically feasible.

The Roadmap

NovaFuse is currently developing a FedEdge proof-of-concept targeting multi-vehicle ISR fusion — where multiple platforms cooperatively track targets using federated sensor fusion models. Each platform maintains a local tracking model trained on its own sensor data. The models federate during fleet-level syncs, producing a global tracking model that benefits from every platform's experience.

We are targeting TRL 4 (laboratory validation of a federated learning pipeline with simulated tactical data) by Q4 2026, with TRL 6 (relevant environment demonstration with two or more live sensor platforms) by Q2 2027.

This timeline targets the TRL 3-5 range — the phase where federated learning transitions from concept to validated system.

The Bottom Line

Federated learning isn't just a machine learning technique. It's an operational paradigm shift. It means your AI gets smarter every time it deploys — not every time someone flies a hard drive back from theatre.

The technology is ready. The operational need is urgent. NovaFuse is building the missing piece: a federated learning framework designed from day one for the constraints of tactical military operations.

Canada has a window to lead in sovereign tactical AI. FedEdge is our bet on how to do it.

NovaFuse Inc. is an Ottawa-based AI company building multi-modal fusion systems for Canadian defence. Federated learning is a core pillar of our edge AI strategy. For partnership inquiries or to learn more about FedEdge, contact info@novafuse.ca.

Learn more about NovaFuse's capabilities:

Research & Publications AI Consulting Services Contact Us