Aggressive Rejection with Adaptive Gradient for Contaminated Data

Contaminated data poses a critical challenge in anomaly detection, as it severely impairs the performance of detection systems. To achieve robustness against contamination, previous approaches have set a specific contamination ratio to manage potential anomalies. However, these methods fail to handle all anomalies in the training dataset even when the assumed ratio aligns with the true contamination level due to the overlap between normal and contaminated data distributions. To address this limitation, we introduce an innovative approach called Aggressive Rejection with Adaptive Gradient (ARAG). ARAG redefines robust training by initially assuming a contamination ratio of 50%, thereby aggressively filtering out probable anomalies. To counterbalance the potential loss of valuable normal data, ARAG dynamically adjusts training gradients based on the distribution of anomaly scores, allowing more representative samples to exert a greater influence on model training. Extensive experiments on three image datasets demonstrate that our strategy outperforms traditional robust loss functions, particularly in high-contamination environments. These results underscore the benefits of our proactive and adaptable approach to managing data contamination and suggest promising new directions for enhancing the resilience of anomaly detection systems in critical applications.