Scholarlink

<p>This paper presents a task-aware semantic model-update compression framework for over-the-air federated learning (OTA-FL) in sixth-generation (6G) edge networks. A coordinate-level relevance score combines loss sensitivity, inter-round innovation, and client representativeness, and jointly governs consensus-skeleton construction, sparsification, unequal-protection quantization, and transmit-power allocation. The edge server combines channel-state information with temporal update correlation for structure-aware aggregate reconstruction. A block-wise candidate-reporting protocol includes both control signaling and OTA payload transmission in the latency model. At the largest evaluated candidate ratio, control signaling accounted for 4.6% of average round latency on KU-HAR and 6.9% on the Chapman-Shaoxing ECG task. For smooth non-convex objectives, the convergence analysis characterizes a stationary neighborhood whose residual separates semantic omission, compression bias, prediction mismatch, and wireless noise. Across 12 independent runs with fading, imperfect channel-state information, synchronization mismatch, non-identically distributed data, and partial participation, the framework achieved <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(93.74\pm 0.31\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>93.74</mn> <mo>±</mo> <mn>0.31</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> accuracy on KU-HAR and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(86.47\pm 0.42\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>86.47</mn> <mo>±</mo> <mn>0.42</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> macro-F1 on ECG. The next-best baseline, top-<i>k</i> magnitude OTA with adaptive quantization, achieved 92.34% and 84.91%, yielding gains of 1.40 and 1.56 percentage points. The corresponding communication times to reach 95% of the dense reference performance were 1.56 and 4.91&#xa0;s, compared with 2.33 and 8.73&#xa0;s for that baseline.</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

Task-aware semantic model-update compression for over-the-air federated learning in 6G edge networks

  • 期刊论文

  • 发表日期：

    2026年7月30日

查看全文
  • Wadhah Al-Zofi

摘要

This paper presents a task-aware semantic model-update compression framework for over-the-air federated learning (OTA-FL) in sixth-generation (6G) edge networks. A coordinate-level relevance score combines loss sensitivity, inter-round innovation, and client representativeness, and jointly governs consensus-skeleton construction, sparsification, unequal-protection quantization, and transmit-power allocation. The edge server combines channel-state information with temporal update correlation for structure-aware aggregate reconstruction. A block-wise candidate-reporting protocol includes both control signaling and OTA payload transmission in the latency model. At the largest evaluated candidate ratio, control signaling accounted for 4.6% of average round latency on KU-HAR and 6.9% on the Chapman-Shaoxing ECG task. For smooth non-convex objectives, the convergence analysis characterizes a stationary neighborhood whose residual separates semantic omission, compression bias, prediction mismatch, and wireless noise. Across 12 independent runs with fading, imperfect channel-state information, synchronization mismatch, non-identically distributed data, and partial participation, the framework achieved \(93.74\pm 0.31\%\) 93.74 ± 0.31 % accuracy on KU-HAR and \(86.47\pm 0.42\%\) 86.47 ± 0.42 % macro-F1 on ECG. The next-best baseline, top-k magnitude OTA with adaptive quantization, achieved 92.34% and 84.91%, yielding gains of 1.40 and 1.56 percentage points. The corresponding communication times to reach 95% of the dense reference performance were 1.56 and 4.91 s, compared with 2.33 and 8.73 s for that baseline.