Quantifying modality imbalance and visual jailbreak robustness in LLaVA via projected gradient descent

Abstract While Large Vision Language Models (LVLMs) exhibit remarkable capabilities, their visual modality introduces a critical attack surface that can bypass text only safety alignments. This paper evaluates the vulnerability of LLaVA-1.5 to targeted adversarial visual prompts designed to induce m...

Popoln opis

Bibliografske podrobnosti

Principais autores:	Saklain Abdullah, Riad Hossain, Mahfuzulhoq Chowdhury
Format:	Artigo
Jezik:	Inglês
Izdano:	Springer 2026-05-01
Serija:	Discover Applied Sciences
Teme:	Vision language model Large language model Projected gradient descent Jailbreak attack
Online dostop:	https://doi.org/10.1007/s42452-026-08793-w
Oznake:	Označite Brez oznak, prvi označite!