Scholarlink

<p>In recent years, One-class deep learning approaches, especially, autoencoder-based, have emerged as leading methods in the field of anomaly detection. However, the high dimensional, disparate, and non-homogenous datasets affect the performance of these models in many applications. To address these concerns, this paper presents an improved Interpolated Implicit Rack-Minimized Deep Support Vector Data Description (IIRMDSVDD) autoencoder. The proposed model focused on optimizing the projected latent space to produce more effective separating hyperspheres, which will significantly improve the precision and robustness of anomaly detection. The IIRMDSVDD combines two main mechanisms into the training process for adopting an improved network architecture. The first mechanism implemented an interpolation regularizer to improve the latent space. The second component extends the concept of minimizing the covariance matrix in the latent space by introducing implicit linear layers that connect the encoder and decoder within the model. To show the effectiveness of the proposed model, extensive experiments have been carried out. The experiments were conducted on both benchmark datasets and Internet of Things (IoT)-specific datasets. On the MNIST dataset, the proposed IIRMDSVDD model achieved an average accuracy of 98.8%, surpassing the second-best method, DASVDD (97.7%), by 1.1%. On the more complex CIFAR-10 dataset, IIRMDSVDD attained an average accuracy of 74.6%, slightly outperforming the second-best method, OCETN (74.5%), by 0.1%. These consistent improvements across datasets validate the effectiveness and robustness of IIRMDSVDD, establishing it as a new state-of-the-art solution for one-class anomaly detection in both traditional and real-world scenarios.</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

An improved interpolated deep SVDD autoencoder with implicit rack minimization

  • 期刊论文

  • 发表日期：

    2026年5月18日

查看全文
  • Abdulkarim M. Katbi,
  • Radhia Elwerghemmi,
  • Riadh Ksantini

摘要

In recent years, One-class deep learning approaches, especially, autoencoder-based, have emerged as leading methods in the field of anomaly detection. However, the high dimensional, disparate, and non-homogenous datasets affect the performance of these models in many applications. To address these concerns, this paper presents an improved Interpolated Implicit Rack-Minimized Deep Support Vector Data Description (IIRMDSVDD) autoencoder. The proposed model focused on optimizing the projected latent space to produce more effective separating hyperspheres, which will significantly improve the precision and robustness of anomaly detection. The IIRMDSVDD combines two main mechanisms into the training process for adopting an improved network architecture. The first mechanism implemented an interpolation regularizer to improve the latent space. The second component extends the concept of minimizing the covariance matrix in the latent space by introducing implicit linear layers that connect the encoder and decoder within the model. To show the effectiveness of the proposed model, extensive experiments have been carried out. The experiments were conducted on both benchmark datasets and Internet of Things (IoT)-specific datasets. On the MNIST dataset, the proposed IIRMDSVDD model achieved an average accuracy of 98.8%, surpassing the second-best method, DASVDD (97.7%), by 1.1%. On the more complex CIFAR-10 dataset, IIRMDSVDD attained an average accuracy of 74.6%, slightly outperforming the second-best method, OCETN (74.5%), by 0.1%. These consistent improvements across datasets validate the effectiveness and robustness of IIRMDSVDD, establishing it as a new state-of-the-art solution for one-class anomaly detection in both traditional and real-world scenarios.