Toward Trustworthy AI: Multi-Target Adversarial Attacks and Robust Defenses for Continuous Data Summarization

This paper presents a study on adversarial attacks targeting continuous data summarization, focusing on multi-resolution image summarization through DR-submodular optimization. It formulates a min-max problem for generating multi-target attacks that degrade multiple summarization models and proposes a robust defense strategy as a regularized max-min problem, accompanied by approximation algorithms with theoretical guarantees. The findings indicate that these adversarial perturbations can significantly impact downstream task performance, highlighting the importance of robust data summarization techniques for maintaining trustworthy AI systems.