Crafting adversarial samples for anomaly detectors in industrial control systems

Ángel Luis Perales Gómez, Lorenzo Fernández Maimó, Alberto Huertas Celdrán, Félix J. García Clemente, Frances Cleary

Research output: Contribution to journalConference articlepeer-review

1 Citation (Scopus)

Abstract

The increasing adoption of the Industry 4.0 paradigm encompasses digitally interconnected factories which enables many advantages. However, it is still necessary to dedicate effort towards investigating protection mechanisms against cyberattacks in these scenarios. Despite the power demonstrated by Anomaly Detection-based Intrusion Detection Systems in industrial scenarios, their vulnerabilities to adversarial attacks, especially to evasion attacks, make Machine Learning and Deep Learning models ineffective for real scenarios. These type of attacks craft samples misclassified by the Intrusion Detection System and potentially reach industrial devices, causing potentially damaging impacts to factory workers and industry resources. Adversarial attacks linked to industrial scenarios are currently in early stages of development, hence most of them have the capability to craft samples misclassified by the IDS but not reach industrial devices. In this work, we present a new adversarial attack named Selective and Iterative Gradient Sign Method that overcomes the limitation of the adversarial attacks present in the literature. To complement this work we also detail a study of how the detection rate of an Intrusion Detection System is degraded and the time required by each technique to generate adversarial samples. The experiments were carried out using a dataset named Electra, collected from an Electric Traction Substation, and showed that adversarial attacks evaluated crafted samples misclassified by the IDS. However, only the method we proposed generated samples that can be understood by intermediate network devices and, therefore, reach their destination. Our experiment outputs demonstrate a lower period of time to achieve and craft adversarial samples using out our iterative based process method as opposed to other current iterative methods currently available.

Original languageEnglish
Pages (from-to)573-580
Number of pages8
JournalProcedia Computer Science
Volume184
DOIs
Publication statusPublished - 2021
Event12th International Conference on Ambient Systems, Networks and Technologies, ANT 2021 / 4th International Conference on Emerging Data and Industry 4.0, EDI40 2021 / Affiliated Workshops - Warsaw, Poland
Duration: 23 Mar 202126 Mar 2021

Keywords

  • Adversarial attacks
  • Anomaly detection
  • Deep learning
  • Industrial control systems

Fingerprint

Dive into the research topics of 'Crafting adversarial samples for anomaly detectors in industrial control systems'. Together they form a unique fingerprint.

Cite this