Scholarlink

<p>Deep learning models are widely adopted in many areas due to their remarkable success in solving different challenging tasks. Despite their great success, adversarial attacks in deep learning manipulate the input, leading to inaccurate predictions. Nowadays, various methods are used to perform formal verification on deep learning. However, these techniques suffer from poor robustness and often suffer from misclassifications in the perturbation bound. In this paper, the deep stacked Autoencoder_Taylor weighted binary softmax (DSA_TaylorWBS) is introduced for adversarial robust learning with abstract interpretation in federated learning framework. The federated learning process is executed in different entities, like node and aggregation server, where the local training is executed in the local data. In training model, the adversarial attack against a network intrusion detection system (NIDS) model is taken under consideration. The input network traffic data is normalized at first, and then augmented using borderline synthetic minority oversampling technique (SMOTE). Following this, the relevant features are selected and the data is finally classified with abstract interpretation using DSA_TaylorWBS. Thereafter, local updation and aggregation are done at the server concerning trust parameters. Further, the DSA_TaylorWBS attained mean square error (MSE) of 0.159, true positive rate (TPR) of 98.888%, and&#xa0;accuracy of 98.279%.</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 trust-aware federated autoencoder framework for adversarial robust learning with abstract interpretation

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Anil Kumar Singh,
  • Abhishek Singh,
  • Manish Gaur,
  • Manik Chandra

摘要

Deep learning models are widely adopted in many areas due to their remarkable success in solving different challenging tasks. Despite their great success, adversarial attacks in deep learning manipulate the input, leading to inaccurate predictions. Nowadays, various methods are used to perform formal verification on deep learning. However, these techniques suffer from poor robustness and often suffer from misclassifications in the perturbation bound. In this paper, the deep stacked Autoencoder_Taylor weighted binary softmax (DSA_TaylorWBS) is introduced for adversarial robust learning with abstract interpretation in federated learning framework. The federated learning process is executed in different entities, like node and aggregation server, where the local training is executed in the local data. In training model, the adversarial attack against a network intrusion detection system (NIDS) model is taken under consideration. The input network traffic data is normalized at first, and then augmented using borderline synthetic minority oversampling technique (SMOTE). Following this, the relevant features are selected and the data is finally classified with abstract interpretation using DSA_TaylorWBS. Thereafter, local updation and aggregation are done at the server concerning trust parameters. Further, the DSA_TaylorWBS attained mean square error (MSE) of 0.159, true positive rate (TPR) of 98.888%, and accuracy of 98.279%.