S$^3$ADNet: Sequential Anomaly Detection with Pessimistic Contrastive Learning

Anomalies are commonly found in sequential data generated by real-world applications, such as cyberattacks in network traffic, human activity changes in wearable sensors. Thanks to the development of computing technology, many impressive results have been obtained from deep learning-based anomaly detection approaches in recent years. This paper proposes a simple neural network framework for detecting anomalies on sequential data, called $S$elf-$S$upervised $S$equential $A$nomaly $D$etection $N$etwork (S$^3$ADNet). S$^3$ADNet first extracts the representations from each data point by performing feature augmentation for contrastive learning; then captures the contextual information from the sequential data points for estimating anomaly probabilities by optimizing the context-adaptive objective. Here, we design a novel loss function based on a pessimistic policy, considering that only anomalies can affect the contextual relationships in sequences. Our proposed method outperformed other state-of-the-art approaches on the benchmark datasets by F1-score with a more straightforward architecture.