A Partially Labeled Anomaly Data Detection Approach Based on Prioritized Deep Reinforcement Learning for Consumer Electronics Security

Abstract

Anomalies within data flows in the Internet of Things environment can potentially result in security vulnerabilities in consumer electronics. Therefore, it is crucial to effectively detect anomaly data to safeguard the reliability and continuous functionality of consumer electronics. Existing related works either learn from unlabeled data using unsupervised methods or leverage the limited labeled data to improve detection performance by semi-supervised methods. However, these methods usually overfit specific types of known anomalies or ignore the uncertainty when model training. To this end, we design a novel approach to jointly optimize the end-to-end detection of labeled and unlabeled anomalies. Specifically, the anomaly data detection problem investigated is first reformulated as a Markov decision process. Then, a partially labeled anomaly data detection approach (PANDA) based on prioritized deep deterministic policy gradient is proposed, which considers uncertainty when the agent makes decisions and can learn from the labeled known anomalies while continuously exploring and detecting prospective anomalies in unlabeled data. Extensive experiments on 13 datasets show that PANDA improves the AUC-ROC and AUC-PR by 3.0%-10.3% and 10.0%-73.5% and its robustness under the impact of anomaly contamination rates compared with four state-of-the-art competing methods.