Researchers have proposed a new type of adversarial examples that are significantly different from the original examples but result in the same answer. They developed algorithms to produce these examples, including NI-FGSM and NI-FGM, which could be used to attack machine learning systems. The study found that these examples are not limited to the neighborhood of the dataset, but are distributed extensively throughout the sample space. This poses security concerns for machine learning models. --- Why it matters: This matters because it shows that current defenses against adversarial attacks may not be effective against this new type of attack, highlighting a need for further research into secure machine learning systems. Source: https://arxiv.org/abs/2510.19347