Classification Attack Detection via Class-specific Visual Context Optimization
Abstract
This final performance report describes three branches of research results performed under the AFOSR award FA2386-20-1-4043. We made significant progress towards state of the art adversarial defense framework for man-in-the-middle (MitM) adversarial attack on neural network classifiers, fast and accurate black-box adversarial attack algorithm for discrete sequential data (e.g. natural languages texts, protein sequences, etc.), and robust policy optimization algorithm for offline reinforcement learning. The research results were disseminated through (i) publications at top publication venues in AI (NeurIPS, AAAI, ICML), (ii) open sourced codes on Github for reproducibility and dissemination, and (iii) lectures for undergraduate and graduate level AI courses at Seoul National University.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 20, 2022
- Accession Number
- AD1194103
Entities
People
- Hyun O Song
Organizations
- Seoul National University