FFAD: Fast Fine-tuning Based Cold-start Anomaly Detection for Serverless Computing Infrastructures

Service auto-scaling is a key feature of serverless computing. The serverless platform is able to automatically create new service instances in nodes or remove existing service instances from nodes according to the automatic schedule. This results in the operational state patterns of infrastructure nodes changing from time to time. Since these pattern changes are often discontinuous as if the algorithm is being applied to a new cold-start environment, the scene is considered a cold-start problem. Existing anomaly detection methods do not specialize in quickly adapting to such discontinuous pattern changes in online environments. Therefore, this paper proposes a cold-start anomaly detection method, FFAD, which is designed for detecting anomalies in multivariate time-series data and can adapt to discontinuous pattern changes through offline pre-training and fast online fine-tuning. For pre-training, a Transformer-based encoder is applied for capturing long-term latent pattern features offline with a proposed iForest-based masking method. For fine-tuning, a specially designed tiny decoder can capture short-term detailed features from the new pattern. After the model is adapted to the current pattern, the encoder-decoder model and a proposed anomaly determination function work together to output a binary classification of anomalies. Extensive experiments prove that FFAD not only achieves fast adaptation to discontinuous changes in node state patterns but also provides an effective anomaly detection capability.