BadAction is a on interactive video generation (IVG) models that poisons the action control channel. By embedding a predefined motion pattern into action sequences and pairing it with a static target video, the model outputs frozen frames that ignore subsequent user actions when triggered. The attack achieves average success rates of 91.0% with action-only triggers and 80.4% with multimodal triggers (image, text, and action together). It also evades existing backdoor detection methods. This is significant because IVG models are used in safety-critical applications like autonomous driving and robotics, and the action interface represents a previously unexplored vulnerability.
Backdoor attacks on interactive video generation make outputs freeze when action triggers appear.
BadAction implants predefined motion patterns into poisoned action sequences, causing triggered models to generate static frames while preserving normal behavior.
Research Lab
Zhihang Wu · Zhongqi Wang · Jie Zhang · Fengming Gu · Shiguang Shan · Xilin Chen
Chinese Academy of Sciences · University of Chinese Academy of Sciences
Research Digest··2 min read
Thread:Agent Security & Attacks
In one line
Backdoor attacks on interactive video generation models can use action-guided triggers to cause frozen frames that ignore user actions.
Before you read this
Two ideas this paper builds onRead it as
The authors propose BadAction, the first backdoor attack tailored to the action control channel of interactive video generation (IVG) models.
The authors systematically study backdoor vulnerabilities in interactive video generation (IVG), a class of models that synthesize action-conditioned videos from an input image and an action sequence.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.