Scholarlink

<p>Multimodal biometric authentication has emerged as a promising solution for improving robustness and security in identity verification systems; however, deploying such models on resource-constrained edge devices remains a significant challenge due to computational, memory and privacy limitations. This paper presents a novel Edge-Native Multimodal Biometric Authentication Framework (EN-MBAF). The novelty lies in the joint co-design of Tiny-Transformer encoders, resource-aware Edge Fusion Transformer and a hybrid Federated Split Learning (FSL) strategy tailored for multimodal biometric systems under strict edge constraints. The proposed architecture utilizes lightweight transformer-based encoders to extract discriminative facial and vocal embeddings while maintaining a minimal memory footprint. Feature-level fusion is performed through the EFT module, which leverages cross-modal attention to learn joint representations efficiently under strict edge constraints. To preserve user privacy and reduce communication overhead, FSL is adopted, enabling decentralized training by sharing only intermediate activations rather than raw biometric data. Experimental evaluation on standard VoxCeleb1 and VoxCeleb2 datasets demonstrates that EN-MBAF achieves an authentication accuracy of 97.85%, an EER of 1.08% and an end-to-end inference latency of 48&#xa0;ms, with a total model size of 4.2&#xa0;MB. Comparative and ablation studies confirm that the proposed framework outperforms existing state-of-the-art methods while offering an optimal trade-off between accuracy, efficiency and privacy, making it well suited for real-world IoT and edge-based biometric applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Tiny-transformer based multimodal biometric authentication with edge fusion and federated split learning

  • 期刊论文

  • 发表日期：

    2026年6月27日

查看全文
  • Khushboo Jha,
  • Aruna Jain

摘要

Multimodal biometric authentication has emerged as a promising solution for improving robustness and security in identity verification systems; however, deploying such models on resource-constrained edge devices remains a significant challenge due to computational, memory and privacy limitations. This paper presents a novel Edge-Native Multimodal Biometric Authentication Framework (EN-MBAF). The novelty lies in the joint co-design of Tiny-Transformer encoders, resource-aware Edge Fusion Transformer and a hybrid Federated Split Learning (FSL) strategy tailored for multimodal biometric systems under strict edge constraints. The proposed architecture utilizes lightweight transformer-based encoders to extract discriminative facial and vocal embeddings while maintaining a minimal memory footprint. Feature-level fusion is performed through the EFT module, which leverages cross-modal attention to learn joint representations efficiently under strict edge constraints. To preserve user privacy and reduce communication overhead, FSL is adopted, enabling decentralized training by sharing only intermediate activations rather than raw biometric data. Experimental evaluation on standard VoxCeleb1 and VoxCeleb2 datasets demonstrates that EN-MBAF achieves an authentication accuracy of 97.85%, an EER of 1.08% and an end-to-end inference latency of 48 ms, with a total model size of 4.2 MB. Comparative and ablation studies confirm that the proposed framework outperforms existing state-of-the-art methods while offering an optimal trade-off between accuracy, efficiency and privacy, making it well suited for real-world IoT and edge-based biometric applications.