Scholarlink

<p>As renewable energy and carbon trading systems converge under the “dual carbon” goals, massive multi-modal data–such as power usage logs, emission metrics, and trading records–must be securely shared across stakeholders. Ensuring the privacy and integrity of such sensitive data is critical for accurate carbon asset accounting and low-carbon grid operations. Federated learning (FL) offers a privacy-preserving framework for collaborative modeling without sharing raw data. This approach is particularly well-suited for power systems, where sensitive and multi-modal data–such as sensor measurements, control logs, and carbon transaction records–must remain local. However, its distributed nature makes it vulnerable to poisoning attacks, especially in heterogeneous and multi-modal environments where data distributions vary widely. To address this, we propose Critical Layer Robust Aggregation (CLRA), a novel FL defense mechanism that assesses the credibility of layer-wise model updates via cosine similarity, detects anomalies using predefined thresholds, and eliminates malicious updates layer by layer. The server then performs adaptive aggregation based on credibility scores. Experiments on three benchmark datasets and a synthetic cross-domain power-carbon dataset demonstrate that <i>CLRA</i> effectively defends against diverse poisoning attacks, including model-poisoning attacks such as LIE, while maintaining model performance. This method proves to be highly effective for secure and reliable training in real-world low-carbon energy systems.</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

A robust federated aggregation algorithm for multimodal data in smart grid scenarios

  • 期刊论文

  • 发表日期：

    2026年1月4日

查看全文
  • Dong He,
  • Jiaxiang Yan,
  • Yanbo Wang,
  • Fei Zhao,
  • Yiwen Xia,
  • Hanxi Li,
  • Wei Wang

摘要

As renewable energy and carbon trading systems converge under the “dual carbon” goals, massive multi-modal data–such as power usage logs, emission metrics, and trading records–must be securely shared across stakeholders. Ensuring the privacy and integrity of such sensitive data is critical for accurate carbon asset accounting and low-carbon grid operations. Federated learning (FL) offers a privacy-preserving framework for collaborative modeling without sharing raw data. This approach is particularly well-suited for power systems, where sensitive and multi-modal data–such as sensor measurements, control logs, and carbon transaction records–must remain local. However, its distributed nature makes it vulnerable to poisoning attacks, especially in heterogeneous and multi-modal environments where data distributions vary widely. To address this, we propose Critical Layer Robust Aggregation (CLRA), a novel FL defense mechanism that assesses the credibility of layer-wise model updates via cosine similarity, detects anomalies using predefined thresholds, and eliminates malicious updates layer by layer. The server then performs adaptive aggregation based on credibility scores. Experiments on three benchmark datasets and a synthetic cross-domain power-carbon dataset demonstrate that CLRA effectively defends against diverse poisoning attacks, including model-poisoning attacks such as LIE, while maintaining model performance. This method proves to be highly effective for secure and reliable training in real-world low-carbon energy systems.