- NIST: AI cybersecurity defense challenges persist.
- Vulnerabilities include poisoning, abuse, privacy, and evasion attacks.
- Unresolved theoretical problems highlight the need for better defenses.
In a recent report titled "Trustworthy and Responsible AI," NIST identifies four types of cyberattacks capable of manipulating the behavior of AI systems and provides key mitigation strategies along with their limitations. The government agency, tasked with developing domestic guidelines for AI model evaluation and red-teaming, reveals that AI systems are increasingly susceptible to attacks from malicious actors capable of evading security measures and triggering data leaks.
NIST computer scientist Apostol Vassilev, one of the report's authors, acknowledges the progress made in AI and machine learning but warns about unresolved theoretical problems in securing AI algorithms. He states, "There are theoretical problems with securing AI algorithms that simply haven’t been solved yet. If anyone says differently, they are selling snake oil."
The report categorizes potential adversarial machine learning attackers into three types: white-box hackers with full knowledge, sandbox hackers with minimal access, and gray-box hackers somewhat informed about an AI system but lacking access to its training data. All three categories pose serious threats, with fraud evolving and presenting new challenges in the AI space.
The NIST report underscores the growing risk as AI becomes more integrated into the connected economy. It highlights that AI systems can malfunction when exposed to untrustworthy data, leading to attacks such as poisoning and abuse. Poisoning attacks involve introducing corrupted data during the training phase, while abuse attacks insert incorrect information into legitimate sources that the AI system absorbs.
The difficulty of making an AI model unlearn a taught behavior adds complexity to AI defense. The report also identifies privacy attacks, aimed at learning sensitive information about the AI or its training data, and evasion attacks, which attempt to alter the system's responses post-deployment.
Despite ongoing efforts, the report emphasizes the lack of foolproof methods for protecting AI and calls for the community to develop better defenses. While acknowledging the challenges, the NIST suggests that adhering to basic cyber hygiene practices can help mitigate potential abuses. The unresolved nature of AI cybersecurity defense signals the need for continued research and collaboration in securing these transformative technologies.
[embedpost slug="/beijing-welcomes-alula-wonder-of-arabia-exhibition/"]
