Testing the robustness and security of network protocol implementations is essential across all domains. We present NetworkFuzzer, a generic, feedback-driven network fuzzer designed to test and analyze protocol implementations by operating directly on real traffic. Unlike traditional code coverage-based fuzzers, NetworkFuzzer works at the network level and employs a closed-loop fuzzing mechanism that dynamically adapts based on server responses. The system incorporates three key components: (i) response-aware fuzzing operators that perform protocol-specific packet mutations, (ii) a Conditional Tabular GAN (CTGAN) model that learns from both normal and abnormal traffic to generate diverse and protocol-compliant test cases, and (iii) Large Language Models (LLMs) that automate the generation of testing rules from protocol specifications. While NetworkFuzzer is protocol-agnostic and applicable to a wide range of network protocols, in this paper we focus on its application to the Digital Imaging and Communications in Medicine (DICOM) protocol, which is commonly used for medical image exchange, to demonstrate its utility in healthcare cybersecurity. Our evaluation shows that NetworkFuzzer effectively executes real-world attacks and generates realistic synthetic traffic, thus enhancing the robustness of testing and training for security systems.