Agree to Disagree: Robust Anomaly Detection with Noisy Labels

Hofmann, Dennis

Etd

Agree to Disagree: Robust Anomaly Detection with Noisy Labels

Pubblico Deposited

Anomaly detection is challenging due to the scarcity of reliable anomaly labels. Recent techniques thus rely on learning from generated lower-quality labels by employing either clean sample selection or label refurbishment to correct the noisy labels. Both approaches struggle with anomaly detection due to conflating anomalous samples with noisy labeled samples. The class imbalance of anomaly detection combined with the asymmetric higher noise rate of anomalies (driven by their high diversity) leads sample selection techniques to unintentionally discard crucial anomaly samples. Label refurbishment methods, relying on anomalies having distinct properties from inliers, such as higher prediction variance, suffer from incorrect label corrections, especially for marginal clean samples. To overcome these limitations, we introduce Unity, a new learning-from-noisy labels approach for anomaly detection that elegantly leverages the merits of both sample selection and label refurbishment. Unity leverages two deep anomaly classifiers to collaboratively select easy samples with clean labels based on prediction agreement and marginal samples with clean labels via disagreement resolution. Thereafter, instead of discarding the remaining samples that may have noisy labels, Unity introduces a feature-space-based metric called ContrastCorr to refurbish these remaining labels. The set of selected and refurbished clean samples are jointly used to robustly update the anomaly classifiers in an iterative label cleaning process. Our experimental study on a rich variety of anomaly detection benchmark datasets demonstrates that Unity consistently outperforms state-of-the-art techniques for learning from noisy labels.

Creator