In order to identify distributed denial-of-service (DDoS) attacks on SIP-VoIP infrastructures in real time, this paper proposes a genetic algorithm-trained modular neural network (MNN) with SMOTETomek balancing. In contrast to traditional firewalls and static filters, our framework leverages modular deep learning and evolutionary optimisation to accurately and adaptively identify malicious traffic. The model incorporates noise reduction and synthetic oversampling using a large-scale SDNDDoS dataset (1,048,575 records) to mitigate the class imbalance problem prior to modular learning. The experimental results from the study demonstrates that the ensemble outperforms similar machine learning and ensemble baselines, achieving 99.5% accuracy with an F1-score of 0.978. Beyond sheer performance, the modular architecture offers parallelized training and robust generalisation, guaranteeing resistance to diverse DDoS vectors. This study demonstrates how well genetic algorithms combined with modular neural networks can be leveraged for intrusion detection, providing a workable and scalable solution to protect next-generation SIP-VoIP communication systems against evolving hostile threats. DDoS Intrusion Detection in SIP-VoIP Networks...