Scholarlink

<p>The proliferation of Generative Artificial Intelligence (GenAI) systems has introduced unprecedented security challenges, with adversarial attacks evolving faster than defensive countermeasures. Golda et al. (IEEE Access 12: 48126–48144, 2024)’s comprehensive survey documented privacy and security concerns across five perspectives including user, ethical, regulatory, technological, and institutional providing valuable awareness of the threat landscape. However, a critical gap still remains: systematically mapping specific attack vectors to their corresponding defense mechanisms to guide practical security implementations. Building upon this foundational work, this study reframes GenAI security through a threat-modeling lens, taxonomizing attack vectors into five primary categories—data poisoning, model inversion, adversarial inputs, inference manipulation, and supply chain attacks—and quantitatively evaluating defense effectiveness against each vector. Through systematic synthesis of the literature, this study constructs a novel attack-defense mapping matrix quantifying thirteen defense mechanisms’ effectiveness across threat categories. This study’s analysis reveals critical protection gaps, particularly against model extraction and deepfake generation. Privacy-preserving techniques like differential privacy and federated learning demonstrate high effectiveness against data poisoning but limited utility against adversarial inputs. This study provides a context-sensitive decision framework enabling security practitioners to select defenses based on threat profiles, resource constraints, and regulatory requirements. Approximately one-third of identified attack vectors lack mature defensive solutions, highlighting priority research areas. This work bridges theoretical security research and practical implementation, providing actionable guidance for securing GenAI deployments across healthcare, finance, and media sectors.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adversarial threat modeling in generative AI: a systematic mapping of attack vectors to defense mechanisms

  • 期刊论文

  • 发表日期：

    2026年5月4日

查看全文
  • Aminu Muhammad Auwal

摘要

The proliferation of Generative Artificial Intelligence (GenAI) systems has introduced unprecedented security challenges, with adversarial attacks evolving faster than defensive countermeasures. Golda et al. (IEEE Access 12: 48126–48144, 2024)’s comprehensive survey documented privacy and security concerns across five perspectives including user, ethical, regulatory, technological, and institutional providing valuable awareness of the threat landscape. However, a critical gap still remains: systematically mapping specific attack vectors to their corresponding defense mechanisms to guide practical security implementations. Building upon this foundational work, this study reframes GenAI security through a threat-modeling lens, taxonomizing attack vectors into five primary categories—data poisoning, model inversion, adversarial inputs, inference manipulation, and supply chain attacks—and quantitatively evaluating defense effectiveness against each vector. Through systematic synthesis of the literature, this study constructs a novel attack-defense mapping matrix quantifying thirteen defense mechanisms’ effectiveness across threat categories. This study’s analysis reveals critical protection gaps, particularly against model extraction and deepfake generation. Privacy-preserving techniques like differential privacy and federated learning demonstrate high effectiveness against data poisoning but limited utility against adversarial inputs. This study provides a context-sensitive decision framework enabling security practitioners to select defenses based on threat profiles, resource constraints, and regulatory requirements. Approximately one-third of identified attack vectors lack mature defensive solutions, highlighting priority research areas. This work bridges theoretical security research and practical implementation, providing actionable guidance for securing GenAI deployments across healthcare, finance, and media sectors.