Year: 2025
Type: article
Abstract: Machine learning models face sophisticated cybersecurity threats from adversarial attacks that exploit fundamental vulnerabilities in AI systems. These attacks include carefully crafted adversarial examples that cause misclassification while appearing normal to humans, model poisoning that introduces backdoors through contaminated training data, and extraction attacks that reverse-engineer proprietary models. Effective defense requires a multi-layered approach combining robust model design techniques such as adversarial training, defensive distillation, and gradient masking with runtime protection strategies, including input sanitization, anomaly detection, and ensemble methods. Organizations must complement these technical measures with rigorous operational protocols, including strict access controls, regular security audits, and comprehensive monitoring. As attackers grow more sophisticated, defense strategies must continually evolve through ongoing collaboration between cybersecurity and AI communities, with promising advances in certifiable robustness and integration with broader security frameworks showing potential for improved resilience.
Author: Gresshma Atluri
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4410249852