Deep support vector data description based on correntropy for few-shot anomaly detection

Qing-Zhen Shang, Jinfu Yang, Fuji Fu, Jiaqi Ma

Abstract: Few-shot anomaly detection aims to identify samples that differ from or are abnormal compared to normal samples using a limited number of training samples. Training deep one-class classifiers requires a large number of normal samples, and it may not even require any anomaly samples. Therefore, deep one-class classification can address the issue of class imbalance in few-shot anomaly detection. However, existing methods based on deep one-class classification lack sufficient utilization of the non-linea r relationships and local feature differences among input samples in high-dimensional feature spaces when addressing class imbalance issues. Moreover, they are sensitive to noise and anomaly samples due to the use of the Euclidean distance loss function. To address these limitations, we propose a Deep support vector data description based on Correntropy for Few-Shot Anomaly Detection (DC-FSAD). Specifically, we introduce an improved loss function that replaces the Euclidean distance loss function in deep one-class classification with correntropy. By utilizing correntropy as a measure of similarity, the new loss function can better capture the non-linear relationships among input samples in high-dimensional feature spaces and fully exploit the local feature differences among samples. Additionally, the width parameter of correntropy can be adaptively adjusted to enhance robustness to noise and anomaly samples. By formulating a new optimization problem and leveraging semi-quadratic optimization techniques, our method achieves a tighter hyper-spherical boundary for accurately describing the distribution of normal samples. Experimental results demonstrate that our proposed method outperforms state-of-the-art methods on benchmark datasets.

External IDs:dblp:journals/dsp/ShangYFM25