Equitable Federated Learning with NCA
Nick Lemke 1,, Mirko Konstantin 0, Henry John Krumb, John Kalkhof, Jonathan Stieber, Anirban Mukhopadhyay · 2025
Read Equitable Federated Learning with NCA by Nick Lemke 1,, Mirko Konstantin 0, Henry John Krumb, John Kalkhof, Jonathan Stieber, Anirban Mukhopadhyay (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-2506-21735.txt.
Introduction
Nick Lemke and co-authors bring shared model training to low-resource clinics in this 2025 paper. Hospitals can train one model together without sending patient data out, a method called federated learning. Clinics in poorer countries often lack strong computers and steady internet.
Their system, FedNCA, uses a light model built for medical image outlining. It can train on cheap edge devices, even common smartphones, and keeps data traffic low. The design is also ready for encryption. The work aims to spread smart imaging tools where trained staff are few. Fair access across sites is a stated goal.
Catalog record
| CREATOR | Nick Lemke 1,, Mirko Konstantin 0, Henry John Krumb, John Kalkhof, Jonathan Stieber, Anirban Mukhopadhyay |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 20 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-21735
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Equitable Federated Learning with NCA Nick Lemke 1,, Mirko Konstantin 0, Henry John Krumb, John Kalkhof, Jonathan Stieber, Anirban Mukhopadhyay Abstract Federated Learning (FL) is enabling collaborative model training across institutions without sharing sensitive patient data. This approach is particularly valuable in low- and middle-income countries (LMICs), where access to trained medical professionals is limited. However, FL adoption in LMICs faces significant barriers, including limited high-performance computing resources and unreliable internet connectivity. To address these challenges, we introduce FedNCA, a novel FL system tailored for medical image segmentation tasks. FedNCA leverages the lightweight Med-NCA architecture, enabling training on low-cost edge devices , such as widely available smartphones, while minimizing communication costs . Additionally, our encryption-ready FedNCA proves to be suitable for compromised network communication . By overcoming infrastructural and security challenges, FedNCA paves the way for inclusive, efficient, lightweight, and encryption-ready medical imaging solutions, fostering equitable healthcare advancements in resource-constrained regions. We make our implementation publicly available at: https://github.com/MECLabTUDA/FedNCA 1 Introduction Federated Learning (FL) is rapidly emerging as a transformative approach in medical imaging [ 1 ] . Unlike traditional machine learning methods that rely on centralized datasets, FL facilitates the collaborative training of models across multiple institutions without the need to share sensitive patient data [ 1 ] . This makes it particularly appealing for applications in healthcare, where adherence to stringent data privacy guidelines, such as HIPAA and GDPR, is paramount [ 2 ] . Even in low-and-middle-income countries (LMICs) with limited resources and a shortage of healthcare professionals, FL has the potential to provide access to high-quality medical AI [ 3 ] . While the benefits are evident, the adoption of FL in LMICs is hindered by several infrastructural and technical challenges [ 4 ] . One of the primary barriers in LMICs is the limited access to high-performance computing resources [ 5 ] , making it difficult to train deep neural networks and implement traditional FL frameworks. Additionally, slow and unreliable internet connections exacerbate the problem, as the exchange of model updates between clients and servers becomes time-intensive and inefficient [ 6 ] . In state-of-the-art FL solutions, models are becoming increasingly larger, with architectures growing more complex [ 7 ] . The rising computational demands necessitate powerful hardware, stable network infrastructures, and substantial energy resources — requirements that are often inaccessible in resource-constrained regions [ 8 ] . As state-of-the-art FL methods evolve, the risk of exacerbating global disparities in AI accessibility increases, underscoring the urgency for lightweight, adaptable, and efficient FL frameworks tailored to LMICs [ 9 ] . Figure 1 : FedNCA a) circumvents problems with low bandwidth internet connections, b) allows for efficient encryption to protect from adversaries or untrusted servers, c) is trainable on diverse hardware including smartphones. To address these challenges, we propose FedNCA, an efficient , lightweight , and encryption-ready FL system for LMICs (Fig. 1 ). Unlike conventional FL systems that require substantial computational resources and high-bandwidth connectivity, FedNCA is optimized for deployment in resource-constrained environments. At its core, FedNCA leverages the Med-NCA backbone [ 10 ] , which is inherently lightweight and efficient. This approach allows training on low-cost edge devices, such as inexpensive smartphones [ 11 ] , which are available even in resource-constrained regions [ 12 ] . Moreover, the NCA architecture has 5000\times fewer parameters than a UNet, thereby lowering communication costs and enhancing accessibility in regions with limited internet bandwidth. This not only significantly cuts operational costs but also democratizes access to cutting-edge medical AI, allowing broader participation in FL without the barriers of expensive hardware and network constraints. Another crucial challenge is the presence of untrusted or even malicious servers [ 13 ] , particularly in rural and remote regions where none of the trusted peers can act as servers. In such scenarios, ensuring data privacy and security becomes a significant concern. Encryption techniques like homomorphic encryption provide strong privacy protection guarantees, even against untrusted or malicious servers [ 14 ] . However, despite its robust security benefits, homomorphic encryption introduces substantial computational overhead [ 15 ] , which can vary depending on the size of the message, potentially limiting its practicality in resource-constrained environments. Due to the lightweight MedNCA architecture, the time required for the encryption and decryption process is reduced by a factor of 1800 compared to state-of-the-art architectures like TransUnet [ 16 ] . By drastically reducing computational overhead, FedNCA provides strong privacy guarantees, making secure FL more accessible to LMICs. Consequently, robust privacy protection is no longer exclusive to those with high computational resources but becomes available in regions that lack powerful infrastructure. To address both, the infrastructure and security challenges associated with FL in LMICs we make the following contributions in this work: 1) We propose a secure and communication-efficient FL algorithm specifically for NCAs, 2) We evaluate the trained models in terms of their segmentation performance and the transmission costs coming up during FL, 3) We study the real-world applicability on smartphones, and 4) we analyze the efficiency improvement of the homomorphic encryption on NCAs. 2 Background Neural Cellular Automata are characterized by their minimal number of parameters, making them highly efficient compared to traditional neural network architectures. Med-NCA [ 10 ] , an NCA-based architecture specifically designed for medical image segmentation, has demonstrated performance comparable to conventional neural networks, like the UNet [ 17 ] while utilizing only a small fraction of the parameters. Med-NCA’s efficiency stems from its iterative application of simple rules. Cellular automata like Conway’s Game of Life are known to be Turing-complete, demonstrating that complex computations can be achieved with inherently simple update rules. NCA-based algorithms adapt this idea and encode the rule in a lightweight neural architecture [ 18 ] , which significantly reduces hardware requirements compared to large architectures. Communication Bottlenecks in FL arise due to its star-shaped topology, where a central server receives updates from multiple clients and then sends aggregated updates back [ 19 ] . These updates typically consist of model parameters or gradients, which are often large in size, leading to significant bandwidth consumption and increased latency [ 6 ] . This challenge is especially pronounced in settings with limited network capacity [ 20 ] . To mitigate this issue, various encoding and compression techniques have been introduced, such as quantization, sparsification, and federated dropout, which aim to reduce the size of updates while maintaining acceptable model performance [ 21 ] . However, these methods come with an inherent trade-off: while they decrease communication costs, they may also introduce a performance drop due to the loss of information in the transmitted updates [ 22 ] . 3 Methodology The deployment of secure FL systems in LMICs is often hindered by high communication overhead and computational demands. FedNCA addresses these challenges by providing an efficient , lightweight , and encryption-ready solution. These key attributes make FedNCA an ideal solution for resource-constrained settings. An outline of our FL algorithm can be seen in Fig. 2 and Alg. 1 . Figure 2 : Our FedNCA setup including a) the two-stage Med-NCA backbone, b) the aggregation of encrypted weights, and c) other clients connected via a weak internet connection. Algorithm 1 FedNCA client and server functions. • \varphi :