RelAD Framework Boosts Relational Data Anomaly Detection

Detecting anomalies in relational databases presents significant challenges due to the high-dimensional and heterogeneous nature of multi-table attributes, as well as the complex connection patterns across foreign-key relationships. Existing anomaly detection methods often fall short in these intricate environments. A new framework, RelAD, has been proposed to tackle these issues. RelAD is a reconstruction-based system that identifies anomalies by analyzing both attribute and relational edge reconstructions. It features a conditional sparse-gated attribute reconstruction module that filters redundant attributes and highlights abnormal semantic blocks. Additionally, RelAD includes a dual-view multi-relational edge reconstruction module that detects relation-specific abnormal connections using both intrinsic and behavioral entity profiles. These attribute and relational signals are then combined through a lightweight fusion module to generate a final anomaly score, demonstrating superior performance and efficiency on new benchmark datasets.