About this document
Task Agnostic Adversarial Attack by Soumadeep Saha is a document available to read on EtoBox.
The document proposes a task agnostic adversarial attack called Mimic and Fool that exploits the non-invertibility of CNN-based feature extractors. It finds adversarial images that are mapped to similar features as the original image by the feature extractor. This ensures the original and adversarial images produce similar outputs for downstream tasks regardless of the specific task or model. The attack achieved success rates of 74.0%, 81.0% and 89.6% on image captioning and VQA models when evaluated. It al
- Author
- Soumadeep Saha
- Language
- EN