Scholarlink

<p>The impact of poisonous multimodal training data in Machine Learning systems such as autonomous vehicles, health care, LLM models etc. leads to serious data accuracy issues. Adding some unnecessary input to training data leads to data pollution or poisoning attacks. These attacks affect machine learning systems and provide wrong results. This paper offers a combined cryptographic approach of Advanced Encryption Standards and RSA algorithms along with bidirectional encoder representations from transformers and vision transformers to prevent such attacks. In the proposed method, AES is used for rapid data encryption and the RSA algorithm is used for key authenticity against common malicious interference or unauthorized entry and modification. In addition, BERT is applied for intelligent analysis of textual data to efficiently verify adverse events of poisoning attempts, whereas ViT helps to analyze the image data appearance to identify poisoning attempts. The proposed methodology is intended to be applied in distributed environments like edge devices. Experimental outcomes also show that the proposed framework provides an efficient protection shield for ML services against poisoning attacks. This helps to secure data management in multimodal systems and also aims at addressing the problem of preventing poisoning attacks. The evaluation measures such as accuracy score, computation time, precision etc. show the effectiveness of this proposed approach.</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

Secure ML: a hybrid defense method to prevent poisoning attacks on machine learning systems

  • 期刊论文

  • 发表日期：

    2026年1月17日

查看全文
  • A. T. Archa,
  • K. Kartheeban

摘要

The impact of poisonous multimodal training data in Machine Learning systems such as autonomous vehicles, health care, LLM models etc. leads to serious data accuracy issues. Adding some unnecessary input to training data leads to data pollution or poisoning attacks. These attacks affect machine learning systems and provide wrong results. This paper offers a combined cryptographic approach of Advanced Encryption Standards and RSA algorithms along with bidirectional encoder representations from transformers and vision transformers to prevent such attacks. In the proposed method, AES is used for rapid data encryption and the RSA algorithm is used for key authenticity against common malicious interference or unauthorized entry and modification. In addition, BERT is applied for intelligent analysis of textual data to efficiently verify adverse events of poisoning attempts, whereas ViT helps to analyze the image data appearance to identify poisoning attempts. The proposed methodology is intended to be applied in distributed environments like edge devices. Experimental outcomes also show that the proposed framework provides an efficient protection shield for ML services against poisoning attacks. This helps to secure data management in multimodal systems and also aims at addressing the problem of preventing poisoning attacks. The evaluation measures such as accuracy score, computation time, precision etc. show the effectiveness of this proposed approach.