Understanding the Effect of Data Augmentation in Self-supervised Anomaly Detection

Self-supervised learning (SSL) has emerged as a promising alternative to create supervisory signals to real-world tasks, avoiding extensive cost of careful labeling. SSL is particularly attractive for unsupervised problems such as anomaly detection (AD), where labeled anomalies are costly to secure, difficult to simulate, or even nonexistent. A large catalog of augmentation functions have been used for SSL-based AD (SSAD), and recent works have observed that the type of augmentation has a significant impact on performance. Motivated by those, this work sets out to put SSAD under a larger lens and carefully investigate the role of data augmentation in AD through extensive experiments on many testbeds. Our main finding is that self-supervision acts as a yet-another model hyperparameter, and should be chosen carefully in regards to the nature of true anomalies in the data. That is, the alignment between the augmentation and the underlying anomaly-generating mechanism is the key for the success of SSAD, and in the lack thereof, SSL can even impair (!) detection performance. Moving beyond proposing another SSAD method, our study contributes to the better understanding of this growing area and lays out new directions for future research.