블록체인 기반 다중 엣지 AI에이전트의 데이터 무결성 검증 모델

초록

If malicious prompts are injected into the tool description during the MCP-based interworking process, or data is tampered with during the transmission process, the AI agent may construct an incorrect execution plan. Although integrity verification can mitigate these vulnerabilities, log tampering attacks against centralized log storage systems render the verification process itself untrustworthy, and can cause severe damage in environments where AI agents automatically collect and process sensor data. Therefore, this paper proposes a method that configures a security middleware to ensure the integrity of sensor data by storing sensor data in IPFS and recording the root CID on a blockchain, enabling AI agents to verify the authenticity of received data. To evaluate the proposed method, experiments were conducted by configuring normal lightweight edge, malicious lightweight edge, secure MCP client, and IPFS. In malicious MCP server detection, accuracies of 100%, 96.5%, and 90.5% were achieved for rug pull, tool poisoning, and tool name conflict attacks, respectively, and log tampering detection during data transmission achieved 100% accuracy. Furthermore, in a sensor data collection environment with 3 second intervals, the security middleware introduced an average latency of 37.54ms, which was confirmed to have a negligible impact on the overall processing flow.

키워드