InstructTA: Instruction-Tuned Targeted Attack for Large Vision-Language Models

Name: InstructTA: Instruction-Tuned Targeted Attack for Large Vision-Language Models
Uploaded: 2026-10-10T05:10:09.445894+00:00
Duration: 480 s
Description: InstructTA is a novel instruction-tuned attack designed to create highly transferable adversarial attacks against large vision-language models (LVLMs). It uses GPT-4 to infer and rephrase instructions, allowing the attack to extract 'instruction-aware features' for precise targeting. This method outperforms existing attacks by optimizing both feature distance and semantic similarity.

InstructTA is a novel instruction-tuned attack designed to create highly transferable adversarial attacks against large vision-language models (LVLMs). It uses GPT-4 to infer and rephrase instructions, allowing the attack to extract 'instruction-aware features' for precise targeting. This method outperforms existing attacks by optimizing both feature distance and semantic similarity.

Open the full episode