Loss-Guided Model Sharing and Local Learning Correction in Decentralized Federated Learning for Crop Disease Classification
Denis MAMBA KABALA, Adel HAFIANE, Laurent BOBELIN, Raphael CANALS · 2025
Read Loss-Guided Model Sharing and Local Learning Correction in Decentralized Federated Learning for Crop Disease Classification by Denis MAMBA KABALA, Adel HAFIANE, Laurent BOBELIN, Raphael CANALS (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-23063.txt.
Introduction
Denis Mamba Kabala and co-authors let farm AI models learn from each other without a central server, in a 2025 paper. Classifying crop diseases from leaf images helps food security, but farms cannot pool their photos freely. Federated learning trains shared models on private data, yet a central coordinator is a single point of failure.
Their decentralized design has each client exchange models directly with peers. A validation loss score guides which models to accept, and the same score corrects local training through an adaptive loss. Experiments on the PlantVillage dataset with three network types showed better accuracy and faster convergence than standard approaches, plus steadier behavior when farms' data differed. The scheme suits agricultural networks with no trusted center.
Catalog record
| CREATOR | Denis MAMBA KABALA, Adel HAFIANE, Laurent BOBELIN, Raphael CANALS |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 70 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-23063
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Loss-Guided Model Sharing and Local Learning Correction in Decentralized Federated Learning for Crop Disease Classification Thanks: This project is funded as part of the MERIAVINO project. MERIAVINO is part of the ERA-NET Cofund ICT-AGRI-FOOD, with funding provided by national sources [ANR France, UEFISCDI Romania, GSRI Greece] and co-funding by the European Union’s Horizon 2020 research and innovation program, Grant Agreement number 862665. Thanks: * Corresponding author’s email: [email protected] (Denis MAMBA KABALA ) Denis MAMBA KABALA, Adel HAFIANE, Laurent BOBELIN, Raphael CANALS Abstract Crop disease detection and classification is a critical challenge in agriculture, with major implications for productivity, food security, and environmental sustainability. While deep learning models such as CNN and ViT have shown excellent performance in classifying plant diseases from images, their large-scale deployment is often limited by data privacy concerns. Federated Learning (FL) addresses this issue, but centralized FL remains vulnerable to single-point failures and scalability limits. In this paper, we introduce a novel Decentralized Federated Learning (DFL) framework that uses validation loss ( L o s s v a l Loss_{val} ) both to guide model sharing between peers and to correct local training via an adaptive loss function controlled by weighting parameter. We conduct extensive experiments using PlantVillage datasets with three deep learning architectures (ResNet50, VGG16, and ViT_B16), analyzing the impact of weighting parameter, the number of shared models, the number of clients, and the use of L o s s v a l Loss_{val} versus L o s s t r a i n Loss_{train} of other clients. Results demonstrate that our DFL approach not only improves accuracy and convergence speed, but also ensures better generalization and robustness across heterogeneous data environments making it particularly well-suited for privacy-preserving agricultural applications. I Introduction Agriculture plays a crucial role in economic, social and environmental development, while ensuring global food security. The sector employs over 27% of the world’s workforce and contributes 4% to global GDP [ 1 ] . However, crop diseases represent a serious threat to agricultural productivity, causing yield losses estimated at between 20% and 40% per year [ 2 ] . These losses compromise not only farmers’ incomes, but also global food security. In addition, untreated plant diseases can introduce toxic substances into the food chain, posing risks to human health [ 3 ] , [ 4 ] . In this context, artificial intelligence (AI) tools stand out as promising solutions for early and accurate disease detection, contributing to sustainable agricultural practices [ 5 ] , [ 6 ] , [ 7 ] , [ 8 ] , [ 9 ] , [ 10 ] . Recent advances in computer vision have made it possible to automate many agricultural tasks, including the detection and classification of crop diseases [ 11 ] , [ 12 ] . Traditional diagnostic approaches, often time-consuming and costly, are gradually being replaced by machine learning (ML) and deep learning (DL) techniques. Advanced models, such as convolutional neural networks (CNNs) [ 13 ] , [ 14 ] , vision transformers (ViTs) [ 15 ] , [ 16 ] , [ 17 ] , or CNN and ViT simultaneously [ 18 ] , [ 19 ] , have proved highly effective in processing complex data. These tools offer high accuracy for classifying crop diseases based on image analysis [ 20 ] , [ 21 ] , thus enhancing agricultural quality and productivity . However, their large-scale deployment is hampered by challenges linked to the collection of consistent data in large quantities and the use of such data, not least because of confidentiality issues. Deep learning models typically rely on large amounts of data, which poses major challenges in terms of privacy, data security and compliance with regulations such as the RGPD. In agriculture, information on production methods, crop disease detection techniques and disease management practices is considered sensitive by economic players, and sharing it represents a competitive risk [ 22 ] , [ 23 ] , [ 24 ] . Federated learning (FL), a machine learning paradigm, offers a viable solution to these challenges. This approach enables several institutions to collaborate in training a global model without the need to share private local data [ 25 ] , [ 26 ] . Only the parameters of the participants’ various local models are sent to a central server to be aggregated to form the global model, which is then returned to all participants [ 27 ] , [ 28 ] . By enabling collaborative model formation without direct sharing of private data with a central server, federated learning reduces data transfer costs while preserving confidentiality [ 29 ] . In addition, by minimizing the sharing of personal data and giving users control over their own data, federated learning reduces the risk of leakage of sensitive data. While centralized FL, with its central server, is widely adopted in various fields [ 30 ] , it remains vulnerable to high communication costs and single point of failure (server) failures. These limitations make the centralized approach ill-suited to agricultural contexts. In this context, decentralized federated learning (DFL) offers a robust alternative by removing dependency on a central server and promoting peer-to-peer communications [ 31 ] , [ 32 ] , [ 33 ] . DFL has established itself as an essential approach to machine learning, enabling collaborative model learning by multiple clients without dependence on a central server [ 34 ] , [ 35 ] . This architecture improves accessibility and resilience, particularly in rural areas with irregular connectivity. In our framework, we have developed an advanced strategy based on validation loss ( Loss_{val} ) to not only guide model sharing between neighbors, but also correct local learning. Model sharing is carried out selectively: each client compares its own Loss_{val} with that of its neighbors. If a neighbor’s Loss_{val} is lower, that neighbor is considered to have a better model, and its model is then shared with the other clients. This ensures that only the best-performing models circulate in the network, improving overall convergence. In addition to sharing, the Loss_{val} is also used to adjust the local learning process. When a customer receives a better model, it records the Loss_{val} associated with that model and incorporates it into the calculation of its local loss function. This approach allows local models to be corrected in line with global performance, while limiting the influence of external models through adaptive control via \lambda . This makes the system both robust and flexible, suitable for a variety of scenarios and environments. In this paper, we present a new approach combining Loss_{val} -based learning sharing and correction in a decentralized federated learning framework applied to crop disease classification. Our experiments, performed on datasets from PlantVillage [ 36 ] , demonstrate that this strategy improves convergence speed, reduces communication costs and increases model accuracy. The remainder of this article is organized as follows: section 2 reviews related work, section 3 presents our method and the tools used, section 4 analyzes experimental results, and section 5 concludes by proposing perspectives for future research. II Related work II-A Detection of crop diseases using deep learning models