Scholarlink

SMS Spam Detection has increasingly garnered attention due to the widespread use of mobile devices. Currently, most SMS spam detection model training methods rely on centralized data collection, which poses numerous privacy threats and creates security vulnerabilities that expose sensitive information. This study aims to propose a training method that does not require data sharing between parties, based on a federated learning system. In this paper, we experiment with FedAvg, FedAvgM, and FedAdam algorithms using a fine-tuned PhoBERT model tailored for the SMS spam classification task. The results show that the FedAvg algorithm achieves high performance with an accuracy of 99.38% in the IID setting, while the FedAdam algorithm proves more effective in the Non-IID setting, yielding a model with an accuracy of up to 98.5%. This study demonstrates that models like PhoBERT trained with FL algorithms can achieve classification capabilities comparable to centralized data training methods, highlighting the significant potential of FL for natural language processing models without the need for centralized data collection.

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

Federated Learning for Vietnamese SMS Spam Detection Using Pre-trained PhoBERT

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Hoang Quang Anh,
  • Pham Tuan Anh,
  • Pham Son Nguyen,
  • Phan Duy Hung

摘要

SMS Spam Detection has increasingly garnered attention due to the widespread use of mobile devices. Currently, most SMS spam detection model training methods rely on centralized data collection, which poses numerous privacy threats and creates security vulnerabilities that expose sensitive information. This study aims to propose a training method that does not require data sharing between parties, based on a federated learning system. In this paper, we experiment with FedAvg, FedAvgM, and FedAdam algorithms using a fine-tuned PhoBERT model tailored for the SMS spam classification task. The results show that the FedAvg algorithm achieves high performance with an accuracy of 99.38% in the IID setting, while the FedAdam algorithm proves more effective in the Non-IID setting, yielding a model with an accuracy of up to 98.5%. This study demonstrates that models like PhoBERT trained with FL algorithms can achieve classification capabilities comparable to centralized data training methods, highlighting the significant potential of FL for natural language processing models without the need for centralized data collection.