ANADOE: Autoencoder-Based Network Anomaly Detection With Outlier Exposure

Anomaly detection (AD) plays a crucial role in cybersecurity, enabling organizations to proactively identify threats and protect sensitive data and resources from malicious activities. AD is widely applied in tasks such as intrusion detection, malware detection, and fraud prevention. Deep Auto-Encoder (DAE) models are commonly used for AD due to their ability to learn normal patterns from benign data in an unsupervised manner, flagging significant deviations as potential threats during inference. However, in real-world scenarios, the assumption of clean, benign training data is often unrealistic, leading to a substantial drop in performance. This paper introduces ANADOE, an AD algorithm designed to train DAEs in the presence of contaminated training datasets. The proposed algorithm, based on outlier exposure, identifies anomalous samples within the training set and optimizes a weighted joint loss function that accounts for both benign and malicious samples. Experimental results on network traffic benchmarks, including network traffic in IoT environment datasets, demonstrate that our approach outperforms baseline methods in contaminated data settings, highlighting its effectiveness for real-world cybersecurity applications.