Why it matters

Multimodal attacks emerging. Understanding shapes multimodal defense.

Advertisement

The architecture

Image with hidden text (OCR).

Adversarial image patches.

Audio watermark payloads.

Multimodal PI vectorsImage + hidden textOCR extractsAdversarial patchsteers modelAudio payloadsubliminal instructionVision-language models newly vulnerable; audio LLMs face similar issues
Multimodal PI.
Advertisement

How it works end to end

Text-in-image steganography.

Adversarial patches: visual perturbations.

Audio: ultrasonic + subliminal.

Defenses: modality-specific filtering.