Abstract
The proliferation of microservices architecture in cloud environments has introduced unprecedented complexity in system monitoring and anomaly detection. Traditional anomaly detection methods often struggle with the sheer volume, velocity, and multimodal nature of log data, particularly when faced with novel, unseen anomalies or logs from newly deployed services, a scenario demanding zero-shot learning capabilities. This paper proposes a novel framework leveraging Contrastive Representation Learning for zero-shot anomaly detection in multimodal log data of cloud microservices. Our approach effectively integrates structured and unstructured log components, extracting rich semantic representations by maximizing agreement between different augmented views of normal log events while pushing dissimilar events apart in a learned embedding space. This unsupervised pre-training enables the model to learn robust, generalizable representations of normal system behavior without requiring labeled anomalies. During inference, any log event deviating significantly from these learned normal clusters is flagged as anomalous, achieving zero-shot detection. Extensive experiments on synthetic microservice log datasets demonstrate that our framework significantly outperforms state-of-the-art baselines in identifying various anomaly types, including those entirely unseen during training, showcasing its efficacy, adaptability, and potential to enhance the reliability and operational efficiency of cloud microservices.
