DAD++: Improved data-free test time adversarial defense

Open Webpage

Gaurav Kumar Nayak, Inder Khatri, Shubham Randive, Ruchit Rawal, Anirban Chakraborty

Published: 01 Jan 2025, Last Modified: 06 Nov 2025Neurocomputing 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•We provide adversarial defense to pretrained models in the absence of training data.•Method has two stages – Adversarial detection and Adversarial correction.•Detection based on source-free domain adaptation, further improved via soft detection.•Correction using human cognition, ignoring high-frequency regions.•Method also works on semi-supervised setup and across different applications.