🤖 AI

Federated Learning for Object Detection: Enabling Collaborative Drone Learning Without Centralizing Data

This paper demonstrates that Federated Learning enables distributed drone fleets to collaboratively train high-performing, privacy-preserving object detection models on local data, achieving performance close to centralized training while significantly outperforming single-drone approaches without the need for data centralization.

Original authors: Daniel M. Jimenez-Gutierrez, Enrique Zuazua, Georgios Kellaris, Joaquin del Rio, Oleksii Sliusarenko, Xabi Uribe-Etxebarria

Published 2026-07-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Daniel M. Jimenez-Gutierrez, Enrique Zuazua, Georgios Kellaris, Joaquin del Rio, Oleksii Sliusarenko, Xabi Uribe-Etxebarria

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a fleet of drones flying over a battlefield, a disaster zone, or a construction site. Their job is to spot important things like tanks, soldiers, or vehicles. To get really good at this, they need to "see" thousands of different examples of these objects.

The Problem: The "All-Your-Photos" Dilemma
Usually, to teach a computer to recognize things, you gather every single photo from every drone and dump them into one giant central server. It's like asking every student in a school to hand over their entire photo album to the principal so he can study them all at once.

But this is a big problem for drones:

Privacy & Security: Some photos might show sensitive locations or private property that can't be shared.
Bandwidth: Sending thousands of high-quality photos from the sky to the ground takes forever and eats up data plans.
Storage: Drones have tiny hard drives; they can't keep everything.

The Solution: The "Study Group" Approach (Federated Learning)
This paper proposes a smarter way called Federated Learning (FL). Instead of sending the photos to the principal, the drones form a "study group."

Here is how it works, using a simple analogy:

The Students (Drones): Each drone keeps its own photos safe in its own pocket. It never sends the photos anywhere.
The Homework (The Model): Each drone tries to learn from its own photos and creates a "study guide" (a set of rules on how to spot a tank).
The Teacher (The Server): The drones send only their study guides to the teacher. They don't send the photos.
The Class Discussion: The teacher mixes all the study guides together to create one "Master Guide" that is smarter than any single student's guide.
The Result: The teacher sends this Master Guide back to all the drones. Now, every drone is smarter because it learned from the group's collective experience, but no one ever saw anyone else's private photos.

What They Did in the Experiment
The researchers from Sherpa.ai tested this idea using a dataset of military drone images called KIIT-MiTA. They compared three scenarios:

The Lone Wolf (Single-drone): One drone tries to learn only from its own limited photos.
The Big Library (Centralized): All photos are sent to a central server to train one giant model.
The Study Group (Federated): Drones train locally and share only their "study guides" (model updates).

The Results: Winning Without Sharing
The results were impressive:

The Lone Wolf was weak: A drone training alone missed a lot of objects because it hadn't seen enough variety.
The Study Group was nearly as good as the Big Library: The Federated Learning approach achieved almost the same high accuracy as the Centralized method (which had all the data).
The Big Gains: Compared to the "Lone Wolf," the Federated approach improved detection accuracy by over 50% to 67%.
Speed: They tested different "lightweight" models (like YOLO26 nano) that are small enough to run on a drone's limited computer. The best one could spot objects very accurately while running fast enough for real-time use.

The "Bad Weather" Test
The researchers also tested what happens if the internet connection is terrible (simulating a weak signal in the field). Even with a very slow connection (600 Kbps), the system still worked well. It just took a bit longer to exchange the "study guides," but it didn't break the system or lower the accuracy.

The Bottom Line
This paper proves that drones can learn to be excellent at spotting objects by working together as a team, without ever having to share their private, sensitive photos. It's a way to get the power of a giant database while keeping every drone's data safe and local.

Technical Summary: Federated Learning for Object Detection in Distributed Drone Fleets

Problem Statement

Object detection is a critical capability for safety-critical drone applications, including disaster response, infrastructure monitoring, and defense. However, achieving robust performance in these dynamic, heterogeneous environments typically requires large, diverse datasets. Conventional approaches rely on centralizing aerial imagery to train models, a practice that faces significant barriers:

Privacy and Regulatory Constraints: Data often involves sensitive geographic information or proprietary operational details, making aggregation legally or ethically problematic.
Operational Limitations: Distributed drone deployments often suffer from limited connectivity, onboard storage restrictions, and high bandwidth costs associated with transmitting large volumes of raw video or image data to central servers.
Generalization Issues: Training on data from a single drone (Single-drone) limits the model's exposure to diverse viewpoints, scales, and environmental conditions, reducing its ability to generalize.

The core challenge addressed is how to leverage the collective knowledge of distributed drone fleets to improve object detection models without violating data sovereignty or incurring prohibitive communication costs.

Methodology

The authors propose and evaluate a Federated Learning (FL) framework specifically for object detection tasks using Horizontal FL (HFL). In this paradigm, multiple nodes (drones) collaboratively train a shared global model while keeping their raw image data local.

Experimental Setup

Dataset: The study utilizes the KIIT-MiTA dataset, a high-resolution drone imagery dataset containing seven military-relevant classes (artillery, missile, radar, multiple rocket launcher, soldier, tank, and vehicle).
Data Partitioning: To simulate a realistic distributed environment, the training data was partitioned across four nodes using a Dirichlet-based protocol ( $\alpha=1$ ). This creates a non-IID (non-independent and identically distributed) setting where each node holds a different distribution of object classes and scenes, mimicking real-world heterogeneity.
Models Evaluated: The study benchmarks four lightweight YOLO nano architectures implemented via the Ultralytics framework: YOLOv5 nano, YOLOv8 nano, YOLO11 nano, and YOLO26 nano.
Training Scenarios: Three distinct scenarios were compared under a common protocol:
1. Centralized: All data pooled at a server for training.
2. Single-drone: Each node trains independently on its local partition only.
3. Federated: Nodes perform local training (10 epochs per round) and exchange model weights (not data) with a central aggregator using Federated Averaging (FedAvg) over 10 communication rounds.
Infrastructure: Federated experiments were executed on AWS EC2 instances (g4dn family) using the Sherpa.ai FL platform.

Evaluation Metrics

Performance was assessed using standard object detection metrics:

mAP@0.50: Mean Average Precision at an Intersection over Union (IoU) threshold of 0.50.
mAP@0.50:0.95: Mean Average Precision averaged over IoU thresholds from 0.50 to 0.95 (stricter localization requirement).
Efficiency: Frames Per Second (FPS) and Giga Floating Point Operations (GFLOPs) were measured to assess suitability for edge deployment.

Key Results

The experiments demonstrate that Federated Learning effectively bridges the gap between isolated local training and centralized training.

1. Performance Gains

vs. Single-drone: Federated training yielded substantial improvements over isolated training. The best model, YOLO26 nano, achieved relative gains of 52.89% in mAP@0.50 and 67.80% in mAP@0.50:0.95 compared to the Single-drone baseline.
vs. Centralized: The Federated models remained competitive with the Centralized baseline. For YOLO26 nano, the Federated mAP@0.50 was 0.7137, compared to 0.7322 for Centralized training, indicating that the FL approach recovers most of the performance benefits of data centralization without the associated privacy risks.
Per-Class Analysis: Improvements were consistent across all seven classes. Classes with lower representation in individual nodes (e.g., Missile, Radar) saw significant gains, confirming that FL helps the model learn robust representations from diverse data distributions.

2. Efficiency and Deployment

Inference Speed: The training strategy (Centralized vs. Federated) did not impact inference speed. All nano models maintained high FPS (approx. 98–106 FPS on GPU) and low computational complexity (6.1 GFLOPs for YOLO26 nano).
Bandwidth Constraints: A stress test simulating limited bandwidth (600 Kbps bidirectional cap) showed that while training time increased (from 2 hours to 3 hours 20 minutes), detection performance remained comparable to the unconstrained baseline. This suggests FL is robust to the communication constraints typical of wireless drone networks.

Significance and Claims

The paper positions its contribution not as a novel algorithm or detector architecture, but as a controlled experimental evaluation of standard FL strategies in a realistic, privacy-constrained drone scenario.

Privacy-Preserving Scalability: The work demonstrates that FL enables scalable, high-performing object detection across distributed fleets without the need to centralize raw aerial imagery, thereby addressing regulatory and operational barriers to data sharing.
Practical Viability: By showing that FL models approach Centralized performance while significantly outperforming Single-drone models, the authors argue that FL is a practical alternative for safety-critical edge AI where data sovereignty is paramount.
Edge Suitability: The results confirm that lightweight models (specifically YOLO26 nano) trained via FL are suitable for deployment on resource-constrained edge infrastructure, maintaining real-time inference capabilities.

The authors conclude that while FL introduces training-time communication costs, it offers a viable trade-off for distributed drone operations, enabling continuous model improvement across diverse operational environments without compromising data privacy or requiring high-bandwidth data transfer.