Authors - Premanand Ghadekar, Utkarsh Patil, Niraj Ukare, Vansh Bhatt, Rohan Uplenchwar, Shreya Sidnale Abstract - Traditional multi-agent communication systems rely on fixed security protocols and static message processing pipelines, leaving them vulnerable to advancing cyberattacks and dependent on expensive infrastructure. This paper introduces a Secure Multi-Agent Communicational Protocol designed as a lightweight, affordable framework for small and medium-scale systems to communicate safely without enterpriselevel costs. The current setup depends heavily on predictable session keys, making systems prone to impersonation, replay attacks, token alterations, and man-in-the-middle interceptions. This framework stimulates agentto- agent interactions through three primary components: a predictive security model, a dual-token authentication mechanism, and a protocolaware attack engine. The infrastructure utilizes WebSocket connections integrated with Redis Pub/Sub for real-time messaging. A dynamic session key generation process works alongside a rotating refresh-token system, ensuring that even if a session key is compromised, attackers still require a valid refresh token. The predictive component features a Protocol-Aware XLNet model with a dual-thread structure to examine message sequences and statistical irregularities. A fusion layer integrates these analyses, reporting a Dual-Thread Consistency Score of 0.87 and a 31% gain in early-warning capability. Experimental evaluations demonstrate 93.5% violation sensitivity, 91.7% replay detection accuracy, and 89.3% attack-type classification accuracy. This approach enables timely identification of replay incidents, interceptions, and protocol tampering. Additionally, an independent XGBoost model filters fraudulent links. These enhancements provide substantial gains in early warning capabilities and consistent classification accuracy across various attack categories.
Open Zoom