4 key takeaways from NIST’s new guide on AI cyber threats>
SC Magazine – Laura French
The National Institute of Standards and Technology (NIST) published a detailed paper titled “Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations,” which highlights strategies and insights regarding cyberattacks targeting machine learning models.
Co-authored by NIST computer scientists, Northeastern University, and Robust Intelligence, Inc., the paper embraces NIST’s Trustworthy and Responsible AI initiative.
Notably, the taxonomy outlined in the guide delves into the three categories of adversarial machine learning (ALM) attacks—white-box, gray-box, and black-box attacks.
Furthermore, the text introduces four key takeaways for cybersecurity professionals, AI developers, and AI tool users:
1\) ALM attacks can be employed with minimal knowledge of models or training data, posing significant threats.
2\) Generative AI, as compared to predictive AI, introduces distinctive risks associated with abuse.
3\) Attackers can utilize remote data poisoning to inject malicious prompts into AI systems.
4\) NIST cautions against the assurance of foolproof methods for safeguarding AI against potential attackers, emphasizing the need for robust protective measures and the potential trade-offs in AI development.
Additionally, the guide underscores the inherent vulnerabilities in AI and machine learning technologies, calling for greater awareness and caution, and highlighting the complexity of securing AI algorithms.
NIST warns against overstating the robustness of AI security solutions and reminds developers of the existing challenges in securing AI systems effectively.
Link: https://www.scmagazine.com/news/4-key-takeaways-from-nists-new-guide-on-ai-cyber-threats