PASTA: A Patch-Agnostic Twofold-Stealthy Backdoor Attack on Vision Transformers PubID: pubid: 0000–0000/00$00.00 © 2026 IEEE

Dazhuang Liu, Yanqi Qiao, Rui Wang, Kaitai Liang, Georgios Smaragdakis · 2026

Read PASTA: A Patch-Agnostic Twofold-Stealthy Backdoor Attack on Vision Transformers PubID: pubid: 0000–0000/00$00.00 © 2026 IEEE by Dazhuang Liu, Yanqi Qiao, Rui Wang, Kaitai Liang, Georgios Smaragdakis (2026) 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-2604-20047.txt.

Description

Liu and four colleagues present PASTA, a patch-agnostic backdoor attack on vision transformers that stays stealthy in both pixel and attention domains. They exploit the Trigger Radiating Effect, in which a patch trigger activates backdoors across neighboring patches through long-range self-attention. Multi-location trigger insertion during training strengthens the effect, and bi-level optimization balances stealth against strength. PASTA hits 99.13 percent attack success across arbitrary patches on average, while improving visual stealth over 144 times, attention stealth about 19 times, and robustness nearly three times.

Catalog record

CREATORDazhuang Liu, Yanqi Qiao, Rui Wang, Kaitai Liang, Georgios Smaragdakis
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE94236
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.20047

Browse the full library · About the Tower of Babel