Download PDFOpen PDF in browserAdversarial Text Generation in Cybersecurity: Exploring the Potential of Synthetic Cyber Threats for Evaluating NLP-based Anomaly Detection SystemsEasyChair Preprint 1305910 pages•Date: April 20, 2024AbstractWith the increasing sophistication of cyber threats, evaluating the robustness of anomaly detection systems has become crucial in ensuring cybersecurity resilience. Traditional evaluation methods often rely on static datasets, which may not adequately capture the diversity and complexity of real-world cyber threats. To address this limitation, this paper explores the potential of adversarial text generation techniques in generating synthetic cyber threats for evaluating the robustness of Natural Language Processing (NLP)-based anomaly detection systems. Adversarial text generation techniques manipulate textual data to create subtle variations that are imperceptible to humans but can potentially deceive NLP-based anomaly detection systems. By leveraging these techniques, synthetic cyber threats can be generated, encompassing a wide range of attack scenarios and evasion strategies. These synthetic threats serve as challenging test cases for evaluating the resilience of NLP-based anomaly detection systems against adversarial attacks. This paper discusses various adversarial text generation methods, including gradient-based approaches, generative models, and evolutionary algorithms, highlighting their strengths and limitations in generating realistic synthetic cyber threats. It also explores the impact of different adversarial perturbations on NLP-based anomaly detection systems, such as synonym substitutions, grammatical alterations, and semantic obfuscation. Keyphrases: Adversarial Text Generation, Cybersecurity, Evasion Strategies, NLP-based Anomaly Detection Systems, Synthetic Cyber Threats, adversarial perturbations, attack scenarios, evaluation, evaluation metrics, robustness
|