Researchers have evaluated FedBISINDO, a privacy-preserving framework for cross-domain sign language recognition designed to operate directly on resource-constrained mobile edge devices under statistical non-independent and identically distributed data conditions. In distributed evaluations, the framework maintained real-time gesture translation capabilities while mitigating common client drift and accuracy degradation issues that typically undermine decentralized learning systems.
The development of automatic sign language recognition systems for inclusive communication has accelerated alongside rapid advancements in artificial intelligence. However, standard centralized architectures rely on continuous raw video transmission from user environments to central servers. This centralized transfer generates substantial communication overhead and introduces heightened privacy risks for participating users who must stream visual recordings of their gestures.
Although federated learning offers an alternative by preserving data privacy on local devices, conventional federated learning approaches face significant operational hurdles. Specifically, standard decentralized setups experience severe performance degradation when exposed to statistical non-independent and identically distributed (non-IID) data distributions among participating clients, as well as cross-domain environmental variations that occur across distinct user recording settings.
To resolve these dual challenges of privacy preservation and cross-domain data variability, the authors proposed the FedBISINDO framework using a dedicated hybrid edge processing pipeline. The pipeline integrates a YOLOv11-Pose spatial engine with a sequential network combining a Convolutional Neural Network and a Bidirectional Long Short-Term Memory network (CNN-BiLSTM) to extract lightweight joint-coordinate representations rather than processing full video streams.
To counter the performance penalties associated with statistical heterogeneity, the system incorporates a trajectory-matching aggregation layer specifically designed to mitigate non-IID client drift. The framework’s recognition performance was analyzed through distributed simulations conducted on 8,010 augmented frames representing 36 distinct vocabulary classes under designated label-skew conditions parameterized at an alpha value of 0.5.
Under these distributed evaluation conditions, the proposed framework achieved a validation accuracy of 87.0 ± 0.8 percent alongside an F1-score of 86.9 ± 0.7 percent. When evaluated against established federated baselines, FedBISINDO demonstrated notable comparative gains, securing an absolute accuracy improvement of 13.4 percent over FedProx and an absolute accuracy improvement of 24.9 percent compared with standard FedAvg algorithms.
Hardware profiling verified the practical viability of the architecture on constrained devices, confirming a compact model footprint consisting of 110,564 parameters and an overall computational complexity of 0.003310 GFLOPs. When profiled on-device across 10 Raspberry Pi edge clients, per-frame inference latency ranged between 12.81 and 23.65 milliseconds, consistently meeting the benchmark constraint of less than 33.33 milliseconds required for real-time mobile sign gesture translation.
Subject of Research: Edge-based federated learning for cross-domain sign language recognition
Article Title: FedBISINDO: edge-based federated cross-domain sign language recognition on non-IID data
Article References: Sampedro, G. A., Cheng, Y., Caballero, A., Putra, M. A. P., Karna, N. B. A., & Fauzi, M. A. (2026). FedBISINDO: edge-based federated cross-domain sign language recognition on non-IID data. Scientific Reports. https://doi.org/10.1038/s41598-026-74283-z
Image Credits: AI Generated
DOI: 10.1038/s41598-026-74283-z
Keywords: FedBISINDO, sign language recognition, federated learning, edge computing, non-IID data, YOLOv11-Pose, CNN-BiLSTM, trajectory-matching aggregation, client drift, Raspberry Pi, cross-domain recognition, joint-coordinate representations, label skew, on-device inference, real-time gesture translation, privacy-preserving AI
Cite Scienmag News
Veronica Carney. (November 22, 2026). Edge-Based Federated Learning Framework Enables Real-Time Sign Language Recognition. Scienmag. https://scienmag.com/edge-based-federated-learning-framework-enables-real-time-sign-language-recognition/
Veronica Carney. "Edge-Based Federated Learning Framework Enables Real-Time Sign Language Recognition." Scienmag, 22 November 2026, https://scienmag.com/edge-based-federated-learning-framework-enables-real-time-sign-language-recognition/. Accessed 22 November 2026.
Veronica Carney. "Edge-Based Federated Learning Framework Enables Real-Time Sign Language Recognition." Scienmag. November 22, 2026. https://scienmag.com/edge-based-federated-learning-framework-enables-real-time-sign-language-recognition/








