Introduces 'Hidden Ads,' a new class of semantic backdoor attacks that inject promotional content into VLM responses based on natural user behavior.
March 31, 2026
Original Paper
Hidden Ads: Behavior Triggered Semantic Backdoors for Advertisement Injection in Vision Language Models
arXiv · 2603.27522
The Takeaway
It demonstrates a highly practical and stealthy threat model where models are compromised to serve ads not by visual triggers, but by the semantic context of user queries (e.g., asking for food recommendations). This shifts the security focus for consumer-facing VLMs from patch-based triggers to behavior-triggered vulnerabilities.
From the abstract
Vision-Language Models (VLMs) are increasingly deployed in consumer applications where users seek recommendations about products, dining, and services. We introduce Hidden Ads, a new class of backdoor attacks that exploit this recommendation-seeking behavior to inject unauthorized advertisements. Unlike traditional pattern-triggered backdoors that rely on artificial triggers such as pixel patches or special tokens, Hidden Ads activates on natural user behaviors: when users upload images containi