Sequential Adversarial Anomaly Detection with Deep Fourier Kernel

We present a novel adversarial detector for the anomalous sequence when there are only one-class training samples. The detector is developed by finding the best detector that can discriminate against the worst-case, which statistically mimics the training sequences. We explicitly capture the dependence in sequential events using the marked point process with a deep Fourier kernel. The detector evaluates a test sequence and compares it with an optimal time-varying threshold, which is also learned from data. Using numerical experiments on simulations and real-world datasets, we demonstrate the superior performance of our proposed method.