Backdoor attacks pose serious security risks in modern network environments, particularly on Windows and Linux-based Operating Systems (OSs) in server systems, often serving as entry points for advanced persistent threats (APTs). Traditional Intrusion Detection and Prevention Systems (IDS/IPS) face challenges in detecting these evolving threats due to their dependence on signature-based detection methods. This paper presents an enhanced Zero Trust Software-Defined Perimeter (ZTSDP) architecture integrated with a machine learning(ML)-enabled Snort Intrusion Detection and Prevention System (SnortML) to address backdoor threats. The ZTSDP framework utilizes dynamic trust evaluation, micro-segmentation, and contextual access controls to reduce the attack surface and restrict lateral movement. SnortML leverages ML models to detect suspicious behaviors and zero-day exploits. The proposed architecture was tested against multiple backdoor attack scenarios, including Remote Access Trojans (RATs), web shells, malware droppers, Boleto ransomware, and the Mirai botnet. A comparative analysis of three configurations—traditional Snort, SnortML without ZTSDP, and ZTSDP with SnortML—demonstrated significant improvements in detection rates and reduced false positives for ZTSDP with SnortML, achieving up to a 95% detection rate during high-intensity attacks. Additionally, the architecture maintains minimal latency, making it suitable for real-time deployment in large-scale server environments. This study emphasizes the importance of combining Zero Trust principles with advanced IDS/IPS technologies to provide adaptive, robust, and context-aware mitigation of backdoor threats.