Systematic review synthesizes backdoor attacks and algorithmic bias in generative models, highlighting intersections between security threats and unfair automated decisions.
Key Points
The survey aims to evaluate the intersection of backdoor security vulnerabilities and fairness issues in generative artificial intelligence systems.
Synthesized state-of-the-art literature covering definitions, evaluation metrics, and methodologies for algorithmic bias, fairness attacks, and backdoor exploits in artificial intelligence.
Reviewed current defense mechanisms designed to mitigate backdoor vulnerabilities and reduce unfair discrimination in machine learning outputs.
Identified commonly used benchmark datasets, existing technical limitations, and open directions for future research.
Demonstrated that adversarial backdoor insertions directly exacerbate model bias, leading to disproportionate and discriminatory decisions against specific groups.
Identified critical gaps in existing defense mechanisms, showing that security countermeasures often fail to simultaneously ensure algorithmic fairness and model integrity.
Established that unifying fairness metrics with adversarial robustness is essential for building safe and ethical generative AI frameworks.