Wednesday 12 March 2025
A team of researchers has made a significant breakthrough in detecting anomalies, a crucial task in many fields such as finance, healthcare, and cybersecurity. The traditional approach to anomaly detection involves combining the scores obtained from different models or algorithms to identify outliers. However, this method can be flawed due to the independence assumption required by most statistical tests.
The new method, developed by researchers, takes into account the correlation between the scores of different models, which is often neglected in current approaches. The team used a novel generalized likelihood ratio test (GLRT) to combine the scores and improve the accuracy of anomaly detection.
The GLRT test is based on a simple yet powerful idea: instead of combining the scores directly, it calculates the log-likelihood ratio between the null hypothesis (the data follows a known distribution) and the alternative hypothesis (the data does not follow the known distribution). This approach allows for more accurate identification of anomalies, even in cases where the traditional method would fail.
The researchers tested their new approach on several datasets, including CIFAR-10, SVHN, LSUN, ImageNet, and CIFAR-100. The results showed that the GLRT test outperformed other methods, such as score combining using classical statistical tests, in detecting anomalies.
In one experiment, the team used a dataset of images from CIFAR-10 (a popular benchmark for image classification) and added novelty images from SVHN (Street View House Numbers). They found that the GLRT test detected the novel images with an accuracy of 97.2%, significantly better than other methods.
Another experiment involved using a dataset of images from SVHN and adding novelty images from CIFAR-10. In this case, the GLRT test detected the novel images with an accuracy of 99.22%.
The researchers also performed an eigen-analysis of the log-GLRT score to gain insight into why their approach worked better than others. They found that the scores of different models were correlated, and dividing them by the corresponding eigenvalues improved performance.
This breakthrough has significant implications for various fields where anomaly detection is crucial. For instance, in finance, it could help identify unusual trading patterns or detect fraudulent activities. In healthcare, it could aid in identifying rare diseases or detecting anomalies in medical test results. In cybersecurity, it could assist in detecting unknown malware or identifying suspicious network activity.
The development of this new approach demonstrates the power of combining scores from different models using a more sophisticated statistical framework.
Cite this article: “Breaking Down Anomaly Detection Barriers with Correlation-Aware GLRT Test”, The Science Archive, 2025.
Anomaly Detection, Glrt Test, Score Combining, Statistical Tests, Likelihood Ratio, Novelty Detection, Image Classification, Deep Learning, Machine Learning, Data Analysis
Reference: Edward T. Reehorst, Philip Schniter, “Score Combining for Contrastive OOD Detection” (2025).
