💻 computer science

FecalFed: Privacy-Preserving Poultry Disease Detection via Federated Learning

This paper introduces FecalFed, a privacy-preserving federated learning framework that utilizes a rigorously deduplicated poultry fecal dataset and adaptive optimization to achieve high-accuracy, non-IID avian disease detection while eliminating the need to centralize sensitive farm data.

Original authors: Tien-Yu Chi

Published 2026-04-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tien-Yu Chi

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 world where farmers can spot deadly bird diseases just by looking at a photo of chicken poop. It sounds gross, but it's actually a brilliant way to save our food supply. However, there's a huge problem: farmers are (rightfully) scared to share their photos because they don't want their competitors or the public to know their flock is sick. Plus, many farms don't have super-computers to run complex AI.

This paper introduces FecalFed, a clever solution that solves the privacy problem and the computer power problem at the same time. Here's how it works, broken down into simple stories and analogies.

1. The "Dirty Laundry" Problem (Data Contamination)

Before building their system, the researchers noticed something weird. They tried to download free chicken poop pictures from the internet to train their AI. But when they looked closely, they found a massive mess.

The Analogy: Imagine trying to learn to cook by reading 100 recipes, but 47 of them are just the same recipe copied, pasted, and resized. If you study those, you think you're an expert, but you're actually just memorizing duplicates.
The Fix: The team cleaned up the internet's "chicken poop library." They found that nearly 47% of the images were just duplicates of the same pictures. They threw out the fakes and resized copies, leaving a clean, unique set of 8,770 images. This is their new, honest "textbook" for the AI.

2. The "Secret Recipe" Problem (Privacy)

Now, imagine 10 different farms. Each farm has a different mix of sick and healthy chickens.

The Old Way: Usually, to train a smart AI, you'd ask all 10 farms to send their photos to one big central computer.
- The Problem: Farm owners say, "No way! If you see my sick chickens, my reputation is ruined, and my business might fail." They keep their data locked in a vault.
The FecalFed Way (Federated Learning): Instead of sending the photos, the AI goes to the farms.
- The Analogy: Think of it like a study group. Instead of everyone bringing their homework to a central library to be graded, the teacher (the AI) sends a "study guide" to each student's house.
- Each student studies their own homework (their own farm's photos) and writes down what they learned (math notes/weights).
- They send only their notes back to the teacher. The teacher never sees the actual homework, just the lessons learned.
- The teacher combines all the notes to create a super-smart "Master Study Guide" and sends it back out. Now everyone is smarter, but no one ever saw anyone else's private homework.

3. The "Uneven Classroom" Problem (Non-IID Data)

In the real world, data is messy. One farm might have 90% sick chickens, while another has 90% healthy ones.

The Problem: If a single farm tries to train its own AI alone, it gets confused. It learns that "poop looks like this" means "sick," but that's only true for that farm. When it tries to guess for other farms, it fails miserably (dropping accuracy to about 65%).
The Solution: By using the "study group" method (Federated Learning), the AI learns from all the different farms. It sees the sick ones, the healthy ones, and the weird ones.
The Result: The AI becomes a master detective. It learned to spot diseases with 90% accuracy, almost as good as if it had seen all the photos in one giant pile (which would be 95%).

4. The "Tiny Brain" Problem (Edge Computing)

Farms often have old phones or small computers, not giant servers.

The Challenge: The smartest AI models are like giant brains—they need too much power to run on a farm.
The Fix: The researchers tested a "tiny brain" version of the AI (called Swin-Tiny).
The Result: Even though this tiny brain is much smaller and lighter, it still got 89.74% accuracy. It's like a brilliant student who can solve complex math problems using just a pencil and a small notebook, rather than needing a supercomputer.

The Big Takeaway

FecalFed is a blueprint for the future of farming. It proves that:

We can clean up the internet's messy data to build better tools.
We can train super-smart AI without ever violating a farmer's privacy.
We can run these smart tools on simple, cheap devices right on the farm.

By keeping data local and sharing only "lessons learned," we can stop bird flu and other diseases before they spread, keeping our food supply safe without asking farmers to give up their secrets.

1. Problem Statement

The paper addresses three critical bottlenecks in deploying AI for poultry disease detection (specifically Highly Pathogenic Avian Influenza, Coccidiosis, Newcastle Disease, and Salmonella):

Data Silos and Privacy: Farms are reluctant to share sensitive health data due to biosecurity risks, commercial competition, and reputation concerns. Centralized training requires uploading raw images, which is often impossible.
Hardware Constraints: Rural agricultural facilities lack the high-bandwidth internet and enterprise-grade compute infrastructure required for massive centralized models.
Data Contamination: Existing open-source agricultural datasets suffer from severe, undocumented data leakage. The authors found that popular public repositories contain massive amounts of duplicate images (resized or re-uploaded), leading to artificially inflated performance metrics and train/test leakage.
Data Heterogeneity (Non-IID): Disease outbreaks are localized. A single farm's data distribution is rarely representative of the global population. Training isolated models on single farms leads to catastrophic overfitting and poor generalization.

2. Methodology

A. Data Curation: `poultry-fecal-fl`

To establish a clean baseline, the authors curated a new dataset called poultry-fecal-fl:

Source: Aggregated 16,513 images from Zenodo and Roboflow repositories.
Deduplication: Implemented a dual-algorithm perceptual hashing pipeline using Average Hash (aHash) and Perceptual Hash (pHash). Images were flagged as duplicates if the Hamming distance for both hashes was $\le 5$ .
Result: This process removed 46.89% of the data (7,743 images), revealing that 77.4% of images in popular synthetic datasets were merely down-sampled duplicates of original field data. The final dataset contains 8,770 unique, leak-free images across four classes.

B. Federated Learning Framework (FecalFed)

The system utilizes a Cross-Silo Federated Learning architecture using the open-source Flower (flwr) framework:

Architecture: Raw images remain local on farm edge devices. Only model weight updates are transmitted to a central server.
Non-IID Simulation: To mimic real-world conditions, the dataset was partitioned across 10 simulated farms using a Dirichlet distribution ( $\alpha = 0.5$ ). This creates highly skewed label distributions (e.g., one farm has mostly Salmonella, another mostly Healthy), simulating localized outbreaks.
Models Evaluated: Four Vision Transformer architectures were tested to balance accuracy and edge deployability:
- Heavy: ViT-B/16 (86M params), Swin-Small (50M params).
- Lightweight: ViT-S/16 (22M params), Swin-Tiny (28M params).
- Training Strategy: Pre-trained backbones were frozen; only classification heads were fine-tuned to minimize memory footprint and communication overhead.
Optimization Strategies:
- FedAvg: Standard data-weighted averaging.
- FedAdam: Adaptive server-side optimization. The server treats aggregated pseudo-gradients as input to an Adam optimizer to stabilize convergence under non-IID conditions.

3. Key Contributions

Clean Benchmark Dataset: Release of poultry-fecal-fl, a rigorously deduplicated dataset that eliminates nearly 47% of contamination found in public repositories, setting a new standard for agricultural AI evaluation.
Proof of Necessity for FL: Demonstration that isolated single-farm training collapses under realistic non-IID conditions, proving that collaborative federated frameworks are essential for robust diagnostics.
Comprehensive Federated Benchmarking: Systematic evaluation of vision transformers under non-IID conditions, showing that FedAdam significantly outperforms standard FedAvg for larger models and provides a stable baseline for edge deployment.

4. Results

Metric	Isolated Single-Farm (Non-IID)	Centralized (Upper Bound)	FecalFed (FedAvg)	FecalFed (FedAdam)
Swin-Small Accuracy	64.86% (±24.95%)	95.10%	89.74%	90.31%
Swin-Tiny Accuracy	64.03% (±24.04%)	93.04%	86.89%	89.74%
ViT-B/16 Accuracy	60.97% (±23.05%)	94.81%	88.77%	90.02%

Performance Recovery: Isolated training failed, dropping accuracy to ~64%. The federated approach recovered performance, with FedAdam + Swin-Small achieving 90.31%, closing the gap to the centralized upper bound to only 4.79%.
Edge Efficiency: The lightweight Swin-Tiny model achieved 89.74% accuracy (within 0.28% of the larger ViT-B/16 baseline) while using less than one-third of the parameters (28M vs. 86M). This establishes it as the optimal candidate for resource-constrained farm devices.
Convergence: Extending training to 20 global rounds further improved performance (up to 91.05% for ViT-B/16) with minimal bandwidth overhead due to the frozen backbone strategy.

5. Significance

Privacy-First Surveillance: FecalFed provides a blueprint for global poultry disease monitoring that respects farm data sovereignty and biosecurity, enabling collaboration without data sharing.
Reproducibility: By exposing and removing massive data contamination in public datasets, the paper highlights a critical reproducibility crisis in agricultural AI and offers a sanitized baseline for future research.
Practical Deployment: The study proves that high-accuracy disease detection is feasible on standard edge hardware (e.g., NVIDIA Jetson Nano, smartphones) using lightweight transformers and adaptive federated optimization, making advanced AI accessible to rural agricultural sectors.