Backdoor Attack Against One-Class Sequential Anomaly Detection Models (BA-OCAD)

BA-OCAD: Backdoor Attack Against One-Class Sequential Anomaly Detection Models

We present BA-OCAD, a clean-label backdoor attack framework targeting one-class sequential anomaly detection models. The attack injects carefully designed triggers into training sequences, allowing adversaries to mislead anomaly detectors during inference while preserving benign performance. Experiments on log datasets demonstrate that BA-OCAD exposes significant vulnerabilities in state-of-the-art models, motivating future research in robust defenses.

Citation

@inproceedings{cheng2024backdoor,
  title={Backdoor attack against one-class sequential anomaly detection models},
  author={Cheng, He and Yuan, Shuhan},
  booktitle={Pacific-Asia Conference on Knowledge Discovery and Data Mining},
  pages={262--274},
  year={2024},
  organization={Springer}
}