Reproduce “Classification-Based Anomaly Detection for General Data” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Liron Bergman, Yedid Hoshen

The paper introduces GOAD (Generalized Open-Set Anomaly Detection), a framework that extends transformation-based self-supervised anomaly detection to general data types, including tabular data, by using random affine transformations. It achieves state-of-the-art accuracy on diverse image and tabular datasets, with an 88.2% ROC-AUC on Cifar10 and 98.4% F1 on KDD, while enhancing robustness against adversarial attacks and training data contamination.

Key insights

  • Random affine transformations provide a universal mechanism for self-supervised anomaly detection, generalizing beyond image-specific geometric transformations to diverse data types.
  • Incorporating an uncertainty constant in open-set classification is crucial for generating reliable anomaly scores when test data is significantly different from the training distribution.
  • The use of randomized transformations offers inherent protection against adversarial attacks by making it difficult to craft examples that transfer across different sets of transformations.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.