MAKING NETWORK TROUBLESHOOTING AGENTS TRAINED USING REINFORCEMENT LEARNING ROBUST TO ADVERSARIAL ATTACKS

Smart overview of the Invention

The patent application discusses a method to enhance the robustness of network troubleshooting agents against adversarial attacks. The approach involves using a language model to address troubleshooting issues in computer networks, guided by user feedback. A device collects a trace of steps taken by the agent to generate answers and re-executes the agent to produce additional responses. Crowdsourced feedback from other users is then gathered to evaluate the consistency of these answers, determining whether the language model should be updated.

Recent advancements in large language models (LLMs) like ChatGPT and GPT-4 have opened new possibilities for various industries. LLMs can interact with tools to perform tasks such as web searching and code execution. Network troubleshooting, a complex task, requires solving simpler problems first, such as identifying network details and rogue access points. Agents trained with reinforcement learning can improve over time through user feedback but are vulnerable to adversarial attacks where malicious feedback degrades performance.

The described method leverages a network of geographically distributed nodes interconnected by communication links, ranging from local area networks (LANs) to wide area networks (WANs). These networks facilitate data exchange between devices using protocols like TCP/IP. The invention is applicable to various network configurations, including private networks, VPNs, and the Internet, ensuring robust communication across diverse environments.

The application also details different site types connected to networks, each with unique configurations. These include single or multiple customer edge (CE) routers connected via primary and backup links, such as MPLS VPN or Internet connections. The invention supports diverse site types, ensuring adaptability to different network setups and service level agreements, enhancing the reliability of network troubleshooting.

Software-defined WAN (SD-WAN) technology is utilized to connect local networks and data centers, allowing virtualized WAN functions and centralized management. This approach supports dynamic routing and monitoring, providing a flexible and efficient network management solution. The invention's techniques are applicable to various network topologies, including high-speed links and data centers, ensuring comprehensive network troubleshooting capabilities.