Abstract

The global requirement for seamless digital communications places a premium on the need for timely Internet congestion detection in relation to preserving the quality of service (QoS) parameters and optimizing resource allocation. The problem with traditional active probing approaches is the overhead and delay associated with them. This research proposes, instead, a framework for real-time congestion detection using passive flow analysis which is better suited for contemporary networking environments owing to its non-intrusive, low-latency nature. The methodology employed consists of flow statistics analysis, particularly over aggregation using inter-arrival packet time intervals, retransmission rates, and throughput drops captured at edge routers or gateway machines within traffic-free environments. The system uses NetFlow or IPFIX logs to extract temporal and structural features which are then processed with a hybrid machine learning model based on decision trees and gradient boosting to detect congestion episodes. Validation of the framework was assessed on real backbone and access network traffic datasets. The results demonstrate highly effective performance with a detection accuracy of 96.3% alongside a 3.1% false positive rate, drastically improving baseline model performance. Additionally, the framework allows for implementation on software-defined networking (SDN) infrastructures providing programmable control responses to congestion events. This study proposes advanced solutions for next-generation networks imposing real-time performance monitoring requirements. The implementation of the proposed system will benefit several internet service providers and enterprise networks by improving overall user experience, enhancing traffic engineering, and strengthening congestion control.