对抗紧密包裹与超球面约束的异常检测模型

Adversarial Compact Wrapping with Hyperspherical Constraint for Anomaly Detection

摘要: 无监督异常检测模型常面临检测边界模糊和泛化能力弱的难题。现有的深度异常检测模型，例如DeepSVDD，虽然通过超球面约束使正常模式特征分布更加紧密，但难以适配不规则的特征分布。此外，传统检测方法在对“紧贴正常边界且具有高度迷惑性的异常分布”进行精准建模方面还存在不足。为解决这些问题，本文提出了一种基于对抗紧密包裹与超球面约束的异常检测模型。该模型融合了紧密包裹学习、对抗学习以及深度超球面约束的思想，其博弈损失函数与现有方法的博弈损失函数完全不同。具体而言，该损失函数综合了正常样本的紧密包裹损失、异常样本的排斥损失以及对抗样本的生成损失，目标是缩小已知类别特征分布区域与检测模型所确定的类别特征分布区域之间的差异。其中，紧密包裹学习有助于使模型的检测边界更加清晰和准确，而对抗学习则使模型能够学习到更加鲁棒的特征表示，深度超球面约束使异常样本与正常样本分隔更大，从而优化模式特征分布建模并增强模型的泛化能力。实验结果表明，在多个数据集上，与多种现有的异常检测模型相比，本文提出的模型表现更为出色。

Abstract: Unsupervised anomaly detection models often encounter the challenges of ambiguous detection boundaries and weak generalization ability. Existing deep anomaly detection models, such as Deep SVDD, although they make the feature distribution of normal patterns more compact through hypersphere constraints, are difficult to adapt to irregular feature distributions. Moreover, traditional detection methods still have deficiencies in precisely modeling “anomalous distributions that closely adhere to the normal boundary and are highly deceptive”. To address these issues, this paper proposes an anomaly detection model based on adversarial tight wrapping and hypersphere constraints. This model integrates the ideas of tight wrapping learning, adversarial learning, and deep hypersphere constraints, and its game loss function is completely different from that of existing methods. Specifically, this loss function combines the tight wrapping loss of normal samples, the repulsion loss of anomalous samples, and the generation loss of adversarial samples, aiming to minimize the difference between the feature distribution region of known categories and the category feature distribution region determined by the detection model. Among them, tight wrapping learning helps to make the detection boundary of the model clearer and more accurate, while adversarial learning enables the model to learn more robust feature representations. Deep hypersphere constraints increase the separation between anomalous and normal samples, thereby optimizing the modeling of pattern feature distributions and enhancing the generalization ability of the model. Experimental results show that on multiple datasets, the proposed model outperforms many existing anomaly detection models.