Scholarlink

<p>Multimodal neural systems have demonstrated strong performance in controlled environments; however, their effectiveness degrades under non-stationary conditions where modality reliability varies due to noise, occlusion, sensor variability, and adversarial interference. Conventional fusion strategies, which rely on static or heuristic weighting, fail to adapt to such dynamic conditions, resulting in suboptimal decision-making and increased error rates. To address this limitation, this paper proposes a reliability-aware attention-based neural fusion framework for adaptive multimodal authentication. The proposed approach integrates modality-specific deep encoders with a trust-driven fusion mechanism that jointly models historical reliability and instantaneous confidence for each modality. A soft gating function enables dynamic modality selection, while a reliability-guided attention mechanism assigns adaptive weights to modality representations based on trust estimates. Furthermore, a closed-loop feedback strategy updates modality reliability over time, allowing the system to continuously adapt under distributional drift. The fusion process is additionally interpreted within a probabilistic framework, establishing connections to expectation-based inference and mixture-of-experts modeling. Extensive experimental evaluation on a multimodal biometric dataset demonstrates that the proposed framework achieves a True Positive Rate of 98.6%, reduces the Equal Error Rate to 1.6%, and attains an Area Under the Curve of 0.991, significantly outperforming unimodal baselines, static fusion, and heuristic adaptive methods. The system maintains robust performance under challenging conditions, including low illumination, acoustic noise, and missing modalities. These results highlight the effectiveness of reliability-aware fusion for robust and adaptive decision-making, providing a generalizable solution for multimodal learning in non-stationary environments. All experiments were conducted under identical preprocessing, training, and evaluation settings to ensure fair comparison and replicability, and the implementation code will be publicly released.</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

Reliability-aware adaptive multimodal fusion framework for robust continuous authentication under non-stationary sensing conditions

  • 期刊论文

  • 发表日期：

    2026年7月22日

查看全文
  • S. C. Rajkumar,
  • D. Yuvasini

摘要

Multimodal neural systems have demonstrated strong performance in controlled environments; however, their effectiveness degrades under non-stationary conditions where modality reliability varies due to noise, occlusion, sensor variability, and adversarial interference. Conventional fusion strategies, which rely on static or heuristic weighting, fail to adapt to such dynamic conditions, resulting in suboptimal decision-making and increased error rates. To address this limitation, this paper proposes a reliability-aware attention-based neural fusion framework for adaptive multimodal authentication. The proposed approach integrates modality-specific deep encoders with a trust-driven fusion mechanism that jointly models historical reliability and instantaneous confidence for each modality. A soft gating function enables dynamic modality selection, while a reliability-guided attention mechanism assigns adaptive weights to modality representations based on trust estimates. Furthermore, a closed-loop feedback strategy updates modality reliability over time, allowing the system to continuously adapt under distributional drift. The fusion process is additionally interpreted within a probabilistic framework, establishing connections to expectation-based inference and mixture-of-experts modeling. Extensive experimental evaluation on a multimodal biometric dataset demonstrates that the proposed framework achieves a True Positive Rate of 98.6%, reduces the Equal Error Rate to 1.6%, and attains an Area Under the Curve of 0.991, significantly outperforming unimodal baselines, static fusion, and heuristic adaptive methods. The system maintains robust performance under challenging conditions, including low illumination, acoustic noise, and missing modalities. These results highlight the effectiveness of reliability-aware fusion for robust and adaptive decision-making, providing a generalizable solution for multimodal learning in non-stationary environments. All experiments were conducted under identical preprocessing, training, and evaluation settings to ensure fair comparison and replicability, and the implementation code will be publicly released.