SafeComm-Agent: A Security-Aware Communication and Inference Optimization Framework for Multi-Agent Embodied Intelligence Systems

Abstract

Multi-agent embodied intelligence systems increasingly integrate vision, language, and sensor-based perception to support collaborative navigation, manipulation, and autonomous decision-making in dynamic environments, but frequent inter-agent communication, redundant large-model inference, and multimodal adversarial attacks can simultaneously increase computational overhead and compromise physical task safety. To address these challenges, this study proposes SafeComm-Agent, a security-aware communication and inference optimization framework that jointly coordinates multimodal attack detection, agent communication, and adaptive reasoning computation. The framework first combines Vision Transformers (ViTs), Vision-Language Models (VLMs), cross-modal consistency analysis, and uncertainty estimation to identify anomalous observations and derive agent-level security risks. It then employs a Graph Neural Network (GNN) to construct a dynamic security-aware communication graph, while Multi-Agent Reinforcement Learning (MARL) determines when to communicate, which agents to contact, and what information to transmit; risk-aware message compression and adaptive Test-Time Compute (TTC) further regulate communication and inference budgets, with Monte Carlo Tree Search (MCTS) providing additional verification for high-risk states. Experiments in Habitat-based simulated embodied environments with controlled multimodal adversarial perturbations show that SafeComm-Agent achieves a task success rate of 94.6% and an attack detection F1-score of 93.2%. Compared with full-communication MARL, it reduces inference cost by 26%, while achieving a 39% reduction in total computational cost compared with full-communication and full-computation execution. These results demonstrate effective joint optimization of security, communication, and adaptive inference.