TCAD-Net: Integrating Transformer and Convolution for Weakly Supervised Anomaly Data Discovery

With the continuous development of consumer electronic device functions and the large-scale data, ensuring device safety and privacy has become increasingly important. Deep learning-based weakly supervised anomaly detection methods have been widely applied in various fields such as consumer finance, significantly improving the accuracy and reliability of discovery. However, existing methods typically either utilize convolution or Transformer, thus unable to effectively leverage both the local features and long-range dependencies of the large-scale samples simultaneously. To address this challenge, this work proposes a novel deep learning-based weakly supervised anomaly detection network, named TCAD-Net, which effectively integrates the advantages of convolution and Transformer architectures, thus possessing efficient anomaly data discovery capability. Specifically, TCAD-Net is a dual-channel framework, where the convolutional shrinkage-based feature extraction branch and the Transformer-based feature extraction branch are used for extracting local features and long-range dependencies of the large-scale samples, respectively. Subsequently, these two branches are merged through a bridging unit and then fed into an anomaly score generator module unit to obtain the anomaly scores of samples. Extensive experiments demonstrate that the proposed TCAD-Net exhibits exceptional performance across various large-scale benchmark datasets. Its average performance in terms of Precision-Recall curve area under the curve and Receiver Operating Characteristic curve area under the curve surpasses that of state-of-the-art methods by 11.74% and 4.43%, respectively. The codes and trained models are released at TCAD-Net.