Learning to Discriminate Adversarial Examples by Sensitivity Inconsistency in IoHT Systems

Area of Science:

• Artificial Intelligence
• Cybersecurity
• Health Informatics

Background:

• Deep neural networks (DNNs) are crucial for analyzing health data in Internet of Health Things (IoHT) systems.
• Adversarial attacks pose a significant threat by introducing malicious adversarial examples (AEs) that compromise DNN performance, particularly with text data like medical records.
• Existing detection methods for textual AEs lack performance and generalizability in IoHT environments.

Purpose of the Study:

• To develop an efficient and structure-free adversarial detection method for DNNs in IoHT systems.
• To address the challenge of detecting AEs in discrete textual representations, even in unknown attack and model-agnostic scenarios.
• To improve the security and reliability of DNN-based health data analysis.

Main Methods:

• Proposed an efficient, structure-free adversarial detection method for text data.
• Leveraged the sensitivity inconsistency between adversarial examples (AEs) and normal examples (NEs) when important words are perturbed.
• Designed an adversarial detector based on extracted adversarial features derived from sensitivity inconsistency.

Main Results:

• The proposed detector achieved high performance, with an adversarial recall of up to 99.7% and an F1-score of up to 97.8%.
• Demonstrated superior generalizability across different attackers, models, and tasks.
• The structure-free detector can be deployed in existing applications without modifying target DNN models.

Conclusions:

• The developed adversarial detection method is effective and efficient for securing DNNs in IoHT systems against text-based attacks.
• The sensitivity inconsistency principle provides a robust foundation for detecting AEs in textual data.
• The method offers a practical solution for enhancing the security and trustworthiness of health informatics systems.