DualTAP: A Dual-Task Adversarial Protector for Mobile MLLM Agents

Researchers propose DualTAP, a privacy protection framework for mobile MLLM agents, which reduces screenshot privacy leakage while maintaining the agent's task success rate. When mobile GUI agents rely on MLLMs, they often send screenshots containing personally identifiable information (PII) to untrusted third-party routers, which can mine this data with their own MLLMs. DualTAP trains a lightweight generator to process screenshots, uses a contrastive attention module to precisely locate PII-sensitive regions, and employs a dual-task adversarial objective to simultaneously minimize the task-preservation loss and the privacy-perturbation loss; the research team also released the PrivScreen dataset of annotated mobile screenshots for dual-task evaluation. Experiments on six MLLMs (including GPT-5) show that DualTAP reduces the average privacy leakage rate by 31.6 percentage points (a 3.0× relative improvement), with a task success rate of 80.8%, only a slight decrease from the 83.6% of the unprotected baseline; the authors state this is the first feasible solution for the privacy-utility trade-off in mobile MLLM agents.