AI can see whatever you want with US engineers' new attack technique

US engineers have developed a novel attack technique called RisingAttacK that can manipulate AI computer vision systems to control what the AI "sees." This method targets widely used vision models in applications such as autonomous vehicles, healthcare, and security, where AI accuracy is critical for safety. RisingAttacK works by identifying key visual features in an image and making minimal, targeted changes to those features, causing the AI to misinterpret or fail to detect objects that remain clearly visible to humans. For example, an AI might recognize a car in one image but fail to do so in a nearly identical altered image. The researchers tested RisingAttacK against four popular vision AI models—ResNet-50, DenseNet-121, ViTB, and DEiT-B—and found it effective in manipulating all of them. The technique highlights vulnerabilities in deep neural networks, particularly in the context of adversarial attacks where input data is subtly altered to deceive AI systems. The team is now exploring the applicability of this