Self-interpreting Adversarial Images
summary
Shows a new attack that embeds imperceptible perturbations into images that act as image soft prompts and steer multimodal LLM outputs to adversary objectives while staying on-topic. Authors train perturbations with synthetic Q&A training so images can force ideological framing, mimic styles like Hemingway or Harry Potter, trigger spam URL injection or alternate languages, transfer across models including some commercial systems, remain effective at small magnitudes, and evade many standard defenses; code is on GitHub.
AI-written summary · about
related talks
28mIndirect Prompt Injection Into LLMs Using Images and Sounds
25mTricking Neural Networks : Explore Adversarial Attacks
21mAdversarial Attacks
36mAdversarial Machine Learning: How Secure Is ML?
46mA Short Introduction to Hacking Artificial Intelligence
21mTricking Neural Networks : Explore Adversarial Attacks
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
50mResponsible AI: Adversarial Attacks on LLMs