![]() Unlike imageĬlassification tasks that have a finite set of class labels, finding visually Taking inspiration from the analysis of neural networks' robustnessĪgainst adversarial perturbations, we propose a novel gray-box algorithm forĬreating adversarial examples in image-to-text models. Which comprises two main components: an image encoder, responsible forĮxtracting image features, and a transformer-based decoder, used for generatingĬaptions. ![]() Download a PDF of the paper titled I See Dead People: Gray-Box Adversarial Attack on Image-To-Text Models, by Raz Lapid and 1 other authors Download PDF Abstract: Modern image-to-text systems typically adopt the encoder-decoder framework,
0 Comments
Leave a Reply. |