CL-MMAD: A Contrastive Learning Based Multimodal Software Runtime Anomaly Detection Method

Abstract

Software plays a critical role in the infrastructure of modern society. Due to the increasing complexity, it suffers runtime reliability issues. Online anomaly detection can detect partial failures within the program based on manifestations exhibited internally or externally before serious failures occur in the software system, thus enabling timely intervention by operation and maintenance staff to avoid serious losses. This paper introduces CL-MMAD, a novel anomaly detection method based on contrastive learning using multimodal data sources. CL-MMAD uses ResNet-18 to learn the comprehensive feature spaces of software running status. MSE loss is used as the objective to guide the training process and is taken as the anomaly score. Empirical results highlight the superiority of MSE loss over InfoNCE loss and demonstrate CL-MMAD’s effectiveness in detecting both functional failures and performance issues, with a greater ability to detect the latter.