Researchers have introduced a novel method called RisingAttacK that effectively manipulates AI computer vision systems, altering what these systems perceive in images. This technique focuses on “adversarial attacks,” where input data is modified to deceive AI into misidentifying key visual features like traffic signals or pedestrians. Such vulnerabilities pose significant risks, especially in areas like autonomous vehicles and healthcare technologies. RisingAttacK employs a systematic approach to identify important visual features within an image and assesses how sensitive an AI system is to changes in those features. This enables attackers to make minimal, targeted alterations, leading to different interpretations by the AI while remaining indistinguishable to human observers. The technique was successfully tested on popular AI models, including ResNet-50 and DenseNet-121. Future efforts will explore defending against these attacks and extending the application of RisingAttacK to other AI systems, such as language models.
Source link
Share
Read more