Year: 2024
Type: article
Abstract: The demand for automating image labelling tasks on large data has increased dramatically in recent years.Instead of compressing and uncompressing images to then perform the necessary tasks, we can train a neural network to perform the task on the compressed image itself.By chaining an image compression convolutional neural network to a semantic segmentation convolutional neural network, we can label images using the compressed representation of images without the need for decoding.However, there are vulnerabilities in these convolutional neural network frameworks that leave them exposed to adversarial attacks.While these attacks are well studied in the pixel domain, the research on their efficacy and ability to be mitigated is not widely explored.In this thesis we present an analysis on how standard adversarial attacks perform on standard compressed-segmentation networks of various rate-distortion parameters.Additionally, we show through our experimental results that we can protect segmentation performance against these attacks with similar techniques to the pixel domain, namely training the dataset with adversarial images as well as non-attacked images.
Author: ROMERO, Robert james
Institution: Waseda University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7020976090