Deep learning models for anomaly detection and robust density estimation

Abstract

As artificial intelligence (AI) systems become more pervasive throughout various fields, the reliability of deep learning models in real-life applications is becoming more important. In this dissertation, a method for improving the robustness of a deep learning model to operate safely within various situations potentially occurring in real life was studied. In the first study of this dissertation, an anomaly detection method is proposed to make the deep learning model operate safely against the abnormal data which is unseen in the training. This study focused on the deep generative model among various deep learning models to detect anomalies. By formulating the anomaly detection problem as a Bayesian hypothesis test, the locally powerful Bayesian hypothesis test using a deep generative model is proposed. In the second study, a robust density estimation method against adversarial examples is proposed to maintain the performance of the deep learning model despite external disturbances. In this study, the flow-based generative model, standing as one of the deep generative models, is extended to a Bayesian flow-based generative model. Compared to the existing models, the Bayesian flow-based generative model has robust performance even in the adversarially generated test data. In the third study, by expanding the second study, a method to improve the generalization ability of deep learning models is proposed to enhance the performance of models in practical applications. In this study, a new prior distribution that can be generally applied to Bayesian deep learning models called ‘inverse reference prior distribution’, is proposed. The inverse reference prior distribution regulates the Fisher information matrix of the Bayesian deep learning models and effectively improves the generalization ability of the deep learning models.