A new report from the U.S. National Institute of Standards and Technology (NIST) highlights the growing privacy and security challenges posed by the increasing deployment of artificial intelligence (AI) systems. While AI technologies have been integrated into online services at a rapid pace, there are significant threats that these systems face throughout different stages of their operations.
NIST warns that potential adversarial manipulation of training data, exploiting vulnerabilities in AI models, and even malicious interactions with these models can lead to the exfiltration of sensitive information. These risks can have severe implications for the availability, integrity, and privacy of AI systems.
One particular concern involves chatbot technologies, such as OpenAI ChatGPT and Google Bard, which have gained popularity. These systems have the potential to generate toxic or false information when prompted with carefully designed language, emphasizing the importance of robust defenses against abuse attacks.
The report categorizes attacks on AI systems into evasion attacks, poisoning attacks, privacy attacks, and abuse attacks. These attacks can be executed by threat actors with varying levels of knowledge about the targeted AI system.
NIST acknowledges the lack of adequate measures to mitigate these risks effectively and urges the tech community at large to develop better defenses. This call for action aligns with recent guidelines released by the U.K., the U.S., and international partners to ensure the development of secure AI systems.
Apostol Vassilev, a computer scientist at NIST, underscores the theoretical challenges in securing AI algorithms that have not yet been resolved. With the potential for significant failures and dire consequences, it is crucial to address these privacy and security risks in AI systems.
As the deployment of AI systems continues to expand, it is imperative for developers, researchers, and policymakers to prioritize the implementation of robust security measures to safeguard against adversarial attacks and protect user privacy.
The source of the article is from the blog regiozottegem.be