7 results
Now showing1 - 7 of 7
- Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , FLUID: federated learning with unlearning and instant drift-recoveryFederated Learning (FL) enables collaborative model training without sharing raw data, but struggles with concept drift, where client data distributions evolve over time. Existing approaches either retain outdated knowledge or suffer from high computational costs due to frequent retraining. This work introduces Federated Learning with Unlearning and Instant Driftrecovery (FLUID), a novel framework that treats concept drift as a selective forgetting problem. By leveraging techniques from Federated Unlearning, FLUID removes obsolete knowledge while adapting to new data patterns. It combines the lightweight unlearning capability of Federated Auxiliary Unlearning (FedAU) with the rapid recovery capability of Rapid Retraining (RRT), without requiring full retraining. Experimental results show that FLUID achieves balanced, all-around performance matching RRT in recovery speed and better convergence, and FedAU in lower accuracy dips during drift, and it outperforms natural recovery capabilities of Federated Averaging (FedAvg), and retraining from scratch.Publicationtype: Conference Paper81 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , Detect, adapt, overcome: mitigating concept drift in federated learningFederated Learning (FL) is a distributed machine learning paradigm where a central server coordinates the training of a global model across multiple decentralized clients. Most FL algorithms assume stationary data-generating processes and neglect concept drift, i.e., the change of data distributions over time. This results in a gradual decline of the model's performance and eventually renders the model obsolete. We investigate the impact of concept drift in FL and propose an active approach to detect and adapt to drift. Our detection method is based on local model loss which allows us to categorize clients based on the type of drift they experience and apply appropriate adaptive measures. Through extensive experiments, we demonstrate the efficiency, accuracy, and precision of our approach: the proposed adaptation approach successfully accelerates convergence for drifted clients, outperforming the baseline algorithm. Even under extreme drift conditions, our global model remains stable. This demonstrates the superior robustness of our approach compared to conventional methods.Publicationtype: Conference Paper61 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , Hierarchical bidirectional aggregation for federated learning under concept driftFederated Learning (FL) enables collaborative model training across decentralized clients while preserving data privacy. However, real-world FL deployments face two fundamental challenges: heterogeneous data distributions and concept drift, where data distributions evolve over time. While traditional FL assumes a star topology, real-world scenarios often employ hierarchical federated learning based on tree topologies that improve scalability and communication efficiency. In this paper, we introduce Hierarchical Bidirectional Averaging (HBA), a novel aggregation strategy for hierarchical federated networks designed to enhance resilience against data heterogeneity and concept drift. Our main result is that HBA significantly strengthens the resilience of FL against concept drift. In particular, our key findings demonstrate that (1) servers higher in the hierarchy exhibit greater accuracy stability and retain generalized model representations while the lower servers capture more localized and fine-grained details, (2) increasing tree height improves overall client accuracy, and (3) our aggregation strategy enhances generalization and facilitates a balanced information flow across the hierarchy, resulting in faster recovery from concept drift.Publicationtype: Conference Paper39 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , Towards WebAssembly-based federated learning(Springer, 2025-02); ; ; WebAssembly is a portable binary instruction format designed to serve as a compilation target for high-level languages. While originally developed to run performance-intensive applications directly in Web browsers, WebAssembly supports these days a number of different hardware platforms across the compute continuum. This makes it a promising option to run services for training and inference in Federated Learning. To the best of our knowledge, there have been only a few practical approaches to realize Federated Learning using WebAssembly. Therefore, in this paper, we present a framework to achieve this. Our prototypical implementation shows that WebAssembly-based Federated Learning applications are highly portable while providing acceptable runtime overhead during model training.Publicationtype: Conference Paper36 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , Machine learning-based positioning using multivariate time series classification for factory environments(Oklahoma International Publishing, 2024-01-25); ; ; Indoor Positioning Systems have gained significance in numerous industrial applications. While state-of-the-art solutions are accurate, their reliance on external infrastructures can lead to considerable costs, deployment complexities, and privacy concerns, making them suboptimal for specific contexts. Recent advancements in machine learning have surfaced as a potential solution, leveraging data solely from onboard IoT sensors. Nonetheless, the optimal machine learning models for IoT's resource constraints remain uncertain. This research introduces an indoor positioning system using motion and ambient sensors tailored for factories and similar settings with predetermined paths. The problem is framed as multivariate time series classification, comparing various ML models. A novel dataset simulating factory assembly lines is utilized for evaluation. Results demonstrate models achieving over 80% accuracy, with 1 Dimensional-Convolutional Neural Networks showing the most balanced performance followed by Multilayer Perceptrons, considering accuracy, memory footprint and latency. Decision Trees exhibit the lowest memory footprint and latency, rendering its potential for practical implementation.Publicationtype: Book part62 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , Machine learning-based positioning using multivariate time series classification for factory environments(2023-08-22); ; ; Indoor Positioning Systems (IPS) gained importance in many industrial applications. State-of-the-art solutions heavily rely on external infrastructures and are subject to potential privacy compromises, external information requirements, and assumptions, that make it unfavorable for environments demanding privacy and prolonged functionality. In certain environments deploying supplementary infrastructures for indoor positioning could be infeasible and expensive. Recent developments in machine learning (ML) offer solutions to address these limitations relying only on the data from onboard sensors of IoT devices. However, it is unclear which model fits best considering the resource constraints of IoT devices. This paper presents a machine learning-based indoor positioning system, using motion and ambient sensors, to localize a moving entity in privacy concerned factory environments. The problem is formulated as a multivariate time series classification (MTSC) and a comparative analysis of different machine learning models is conducted in order to address it. We introduce a novel time series dataset emulating the assembly lines of a factory. This dataset is utilized to assess and compare the selected models in terms of accuracy, memory footprint and inference speed. The results illustrate that all evaluated models can achieve accuracies above 80 %. CNN-1D shows the most balanced performance, followed by MLP. DT was found to have the lowest memory footprint and inference latency, indicating its potential for a deployment in real-world scenarios.Publicationtype: Preprint34 224 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , DiNS: Nature Disaster in Network Simulations(IEEE, 2022-12); ; ; ; Wireless sensor networks (WSNs) is a promising solution for disaster management because of its scalability and low cost operations. However, testing the effectiveness of WSNs in real-world disasters is time consuming, costly and in some cases even infeasible. In this paper, we propose a Disaster in Network Simulations (DiNS) framework based on OMNeT ++ to replicate a WSN deployed in a disaster in a simulated environment. DiNS allows researchers to observe how the disaster influences the sensor network and how the network can respond in return. A generic coupling interface is developed to support different disaster types. Moreover, we develop a verification tool for functional debugging and verification during new functions de-veloping of sensor nodes, and an optimization tool to support the mathematical optimization of the network operations in response to the disaster. The functionality of DiNS is demonstrated with a case study using wildfire disaster. It provides an easy way for validating and optimizing the disaster management with WSNs.Publicationtype: Conference Paper30