Query-Efficient Video Adversarial Attack with Stylized Logo on Service Computing
Researchers have developed a novel black-box video adversarial attack framework called Stylized Logo Attack (SLA). This method aims to generate more natural and effective adversarial examples for video classification systems using Deep Neural Networks (DNNs). SLA utilizes a three-stage process involving a style reference set for logos, reinforcement learning for optimal logo placement and style, and step-by-step perturbation optimization to enhance fooling rates. AI