Bridging the Task Gap: Multi-Task Adversarial Transferability in CLIP and Its Derivatives

Kuanrong Liu, Siyuan Liang, Cheng Qian, Ming Zhang, Xiaochun Cao · 2025

Read Bridging the Task Gap: Multi-Task Adversarial Transferability in CLIP and Its Derivatives by Kuanrong Liu, Siyuan Liang, Cheng Qian, Ming Zhang, Xiaochun Cao (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-23917.txt.

Description

An attack crafted for one vision-language task can wound a model at another, and the traffic runs both ways. Kuanrong Liu and 4 co-authors map that cross-task transfer for CLIP and its descendants across retrieval, detection, and segmentation. Attacks born in fine-grained tasks like detection and segmentation transfer more strongly than those from coarse tasks, and they hit the original CLIP hard.

Their framework, MT-AdvCLIP, adds a task-aware feature aggregation loss to craft perturbations with wider cross-task reach. On public data sets it raised average attack success across tasks by over thirty-nine percent at the same perturbation budget. The study frames multi-task robustness as its own security question.

Catalog record

CREATORKuanrong Liu, Siyuan Liang, Cheng Qian, Ming Zhang, Xiaochun Cao
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE32 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-23917

Browse the full library · About the Tower of Babel