Perception Is Moving to the Edge
For the last decade, progress in computer vision has followed a simple recipe: collect more images, pool them in one place, train a bigger model.
That recipe is breaking down exactly where perception matters most.
Drones and edge-vision systems now operate in disaster response, infrastructure monitoring, operational security environments, and defense. In all of them, the data that would make models better is generated onboard, in the field, under conditions that make transferring it to a central server impractical, expensive, or simply not permitted.
The images exist. The bandwidth, the permissions, and the governance framework to move them often do not.
The paper “Federated Learning for Object Detection: Enabling Collaborative Drone Learning Without Centralizing Data” shows what happens when you stop trying to move the data and move the training instead.
Figure 1: Global map illustrating drone mAP@0.50 for the best federated model. Individual drones reach between 33% and 44%. The federated model reaches 71%.
Every Drone Sees a Partial World
Aerial object detection is harder than ground-level detection, and for structural reasons.
Objects are small. Viewpoints shift constantly. Scale changes with altitude. Backgrounds are cluttered. The camera itself is moving.
Robust performance under these conditions requires training data that covers a wide range of viewpoints, object scales, sensors, weather conditions, and mission profiles.
In practice, that diversity is never in one place. It is distributed across different drones, different operators, and different deployment sites.
Each node ends up holding a partial view of reality:
- A limited set of environments
- A limited set of target types
- A limited set of operating conditions
- A limited set of edge cases
A model trained only on that partial view learns a partial world. It performs acceptably on what it has seen and poorly on everything else.
Why the Imagery Cannot Simply Be Pooled
The conventional answer is to centralize everything.
In drone deployments, that answer collides with four separate constraints at once.
Privacy and regulation, where aerial imagery may capture sensitive locations, infrastructure, or people. Connectivity, where field operations rarely have the bandwidth to upload continuous high-resolution imagery. Onboard storage, which is finite. And data governance, where imagery may belong to different operators, agencies, or jurisdictions that cannot legally or contractually merge it.
The bottleneck is not the detector architecture.
It is the assumption that all images can end up in the same repository.
Figure 2: Classical centralized training architecture, where private data from every drone must be exchanged and pooled. (Adapted from Figure 4 of the paper)
Federated Learning Changes What Has to Travel
Federated Learning inverts the flow.
Instead of moving images to the model, the model moves to the images.
Each drone or node trains locally on its own data. After a fixed number of local epochs, only the model weights are sent to the Sherpa.ai platform, which aggregates them into an improved global model and redistributes it. The cycle repeats until convergence.
No raw imagery is transmitted at any point in the process.
- Data never leaves the node
- No central image repository is required
- No operator gives up ownership of its footage
- Every node ends up with the improved global model
Figure 3: Federated training architecture on the Sherpa.ai platform. Learnings are exchanged; private data is not. (Adapted from Figure 5 of the paper)
A Deliberately Realistic Experiment
The study uses KIIT-MiTA, a drone-imagery dataset for military object detection with seven target classes: artillery, missile, radar, multiple rocket launcher, soldier, tank, and vehicle.
The training data is split across four nodes using a Dirichlet partition, which produces genuinely uneven class distributions between drones. This is the situation that actually occurs in the field, where different drones cover different areas and encounter different targets.
Three scenarios are then compared under a single, common protocol:
- Single-drone: each node trains alone on its own partition
- Centralized: all data pooled on one server
- Federated: collaborative training via FedAvg on the Sherpa.ai platform, with no raw data exchange
Four lightweight detectors were evaluated in each scenario (YOLOv5 nano, YOLOv8 nano, YOLO11 nano and YOLO26 nano), and every configuration was repeated across 10 independent runs with different random seeds.
Collaboration Recovers Almost All of the Lost Accuracy
The results are consistent across every model tested.
Federated training improves substantially over isolated single-drone training, and lands close to centralized training.
For the best model, YOLO26 nano:
- Single-drone: 0.4668 mAP@0.50
- Federated: 0.7137 mAP@0.50
- Centralized: 0.7322 mAP@0.50
That is a relative improvement of 52.89% over training alone, and 67.80% on the stricter mAP@0.50:0.95 metric, while keeping every image where it was captured.
The gap that matters here is not the small one between federated and centralized.
It is the large one between federated and going it alone.
Figure 5: Single-drone vs Centralized vs Federated for YOLO26 nano. (Adapted from Figure 8 of the paper)
The Benefit Reaches Every Target Class
Aggregate metrics can hide uneven behaviour, so the study also breaks results down class by class.
Federated training improves over the single-drone baseline for all seven classes and on both metrics.
The largest gains appear precisely where they should: on classes such as missile, radar, artillery, and vehicle, where any individual drone simply has too few examples to learn a robust representation.
In several classes, the federated model even edges slightly ahead of the centralized one at mAP@0.50.
Figure 6: Per-class mAP@0.50 across the three training scenarios. (Adapted from Figure 9 of the paper)
Nothing Changes at Inference Time
For edge deployment, accuracy is only half the question. The other half is whether the model can still run on the hardware available.
It can.
Inference throughput was effectively identical across all three training scenarios, which means the federated approach introduces no inference-time overhead whatsoever. Once training finishes, each drone runs the detector locally with exactly the same computational profile as a centrally trained model.
The federated YOLO26 nano model sustains 98.6 FPS at 6.1 GFLOPs, the lowest computational cost of the four architectures evaluated.
The strongest model in the study is also the cheapest one to run
And It Works at 600 Kbps
Federated Learning shifts the cost, from moving data to exchanging parameters during training. So the study stress-tested that cost directly.
The experiment was re-run using YOLO11 nano, the smallest variant evaluated and therefore the most realistic candidate for a constrained drone setup, under a bidirectional cap of 600 Kbps per node enforced with a traffic shaper, and on CPU only.
Detection performance remained comparable to the full-bandwidth baseline. Limited connectivity had a negligible impact on model quality.
The cost appeared in training time, and it was modest, with roughly 16 MB of aggregated parameters uploaded and downloaded by each node per round:
- Federated, unconstrained: 2 hours
- Federated, capped at 600 Kbps: 3 hours 20 minutes
- Centralized: 4 hours 30 minutes
Both federated configurations were faster than centralized training, and neither required moving a single image.
Figure 7: Training time across centralized, federated, and bandwidth-constrained federated scenarios. (Adapted from Figure 10 of the paper)
What Federated Learning Does, and What It Does Not
Avoiding raw data exchange is a substantial reduction in exposure. It is not, on its own, a complete privacy guarantee.
Model updates can still leak information about local data under certain classes of attack. For high-security and operationally sensitive deployments, Federated Learning should be combined with additional protections such as secure aggregation or differential privacy.
Being precise about this is not a caveat. It is the difference between a privacy architecture and a privacy claim.
What This Changes for Drone Operators
Federated Learning allows a fleet, a coalition, or a set of independent operators to build a shared perception capability while each participant retains full control and full ownership of its own data.
- Better generalisation across environments and conditions
- Continuous improvement as new data is captured in the field
- Collaboration across operators, agencies, or jurisdictions
- Full data locality and sovereignty
- Real-time inference preserved on constrained edge hardware
Data Stays Local. Perception Becomes Collective.
What this work demonstrates is not a new detector. It is a different operating model for distributed perception.
Every drone keeps its own images.
Every drone benefits from what the rest of the fleet has learned.
The intelligence becomes shared without the data ever being shared.
For any organisation operating distributed sensing under privacy, bandwidth, or sovereignty constraints, that distinction is the whole point.