A recent report by the National Institute of Standards and Technology (NIST) sheds light on potential attacks that could compromise the security of artificial intelligence (AI) systems, while also offering strategies to mitigate these risks. Titled “Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations,” the report highlights various types of attacks, including evasion, poisoning, privacy breaches, and abuse. NIST categorizes these attacks based on the attacker’s knowledge, goals, objectives, and capabilities.
According to Alina Oprea, one of the co-authors of the report, many of these attacks can be easily executed with minimal knowledge of the AI system and limited adversarial capabilities. For example, poisoning attacks can be carried out by manipulating a few dozen training samples, which constitute a small percentage of the entire training set.
To mitigate the risks associated with poisoning attacks, the report suggests several approaches. These include sanitizing data, modifying the machine learning training algorithm, and implementing robust training methods instead of traditional training protocols. By adopting these strategies, developers and users of AI systems can significantly enhance the security and reliability of their applications.
The release of this report is part of NIST’s ongoing efforts to promote the development of trustworthy AI. It is essential to address the potential vulnerabilities in AI systems to ensure their effectiveness and integrity in various domains.
To learn more about the latest advancements and challenges in AI, industry leaders and federal officials can register for the Potomac Officers Club’s 5th Annual Artificial Intelligence Summit scheduled for March 21st. The summit will provide a platform for meaningful discussions on the latest developments and best practices in securing AI systems.
The source of the article is from the blog meltyfan.es