Demo: Detecting Data-Leak Using NVIDIA Morpheus

Abstract

Translate Article

As Artificial Intelligence (AI) adoption accelerates across industries, the need for robust real-time cybersecurity solutions becomes increasingly critical. Traditional security frameworks that rely on static rules, signature-based detection, and threshold alerts struggle to detect sophisticated identity-based attacks that exploit behavioral anomalies rather than explicit policy violations. These limitations, combined with the growing volume of telemetry data and the latency of conventional detection methods, leave organizations vulnerable to prolonged breaches and data exfiltration. This paper presents a use case with a scalable, AI-driven approach to sensitive data leak detection using NVIDIA Morpheus, a GPU-accelerated cybersecurity framework. We demonstrate how Morpheus, in conjunction with NVIDIA BlueField DPUs and DOCA, enables high-throughput, low-latency inspection of real-time network telemetry without impacting system performance. The solution leverages a pre-trained NLP-based Sensitive Information Detection model capable of identifying ten categories of sensitive data within packet payloads, including credentials, PII, and cryptographic keys. Network-captured PCAP data serialized in JSONlines format passes through a linear, feed-forward Morpheus pipeline, configured via CLI and powered by Triton Inference Server with dynamic batching for optimized inference. The final output provides binary classification results that indicate the presence or absence of sensitive information. This case study illustrates how AI-enhanced telemetry analysis can significantly reduce detection and response times, offering a proactive defense mechanism against modern cyber threats.