The US Department of Commerce’s National Institute of Standards and Technology (NIST) has released a new open source software package, Dioptra, that allows developers to determine what type of attacks would make an AI model perform less effectively. “Testing the effects of adversarial attacks on machine learning models is one of the goals of Dioptra, a new software package aimed at helping AI developers and customers determine how well their AI software stands up to a variety of adversarial attacks,” the NIST said in a statement. The software package, available for free download, can also help developers of AI systems quantify the performance reduction of a model so that they can learn how often and under what circumstances the system would fail, the NIST explained. The release of Dioptra is linked to President Biden’s executive order passed in 2023 that required the NIST to help with model testing. Along with the new software package, the NIST has also released several documents promoting AI safety and standards in line with the executive order. One among those documents is the initial public draft of its guidelines for developing foundation models, dubbed Managing Misuse Risk for Dual-Use Foundation Models. The guidelines outline voluntary practices developers can adopt while designing and building their model to protect it against being misused to cause deliberate harm to individuals, public safety, and national security. The draft offers seven key approaches for mitigating the risks that models will be misused, along with recommendations on how to implement them and how to be transparent about their implementation. “Together, these practices can help prevent models from enabling harm through activities like developing biological weapons, carrying out offensive cyber operations, and generating child sexual abuse material and nonconsensual intimate imagery,” the NIST said, adding that it was accepting comments on the draft till September 9.

Full report : NIST’s Dioptra, an open source AI tool allows developers to determine what type of attacks would make the model perform less effectively.