Towards the Detection of Unobservable Losses in Real-Time Communications

Packet loss, an omnipresent issue that degrades the QoE in Real-time Transport Protocol (RTP)-based real-time communications (RTC) applications, serves as a pivotal indicator for gauging network performance. Conventionally, loss detection hinges on sequence number irregularities. However, many contemporary applications incorporate customized mechanisms that diverge from the standard, confounding loss identification. Although the actual losses are transparent to applications themselves, they remain unobservable to other entities such as network operators, hampering the prospect of overall network management and performance optimization. To address this challenge, we investigate multitudinous RTC traffic gathered across various locations and times. Consequently, we uncover two types of anomalous patterns pertaining to sequence numbers. To discern between factual losses and aberrations in RTP flows, i.e., to detect the unobservable losses, we curate three distinct datasets, aggregating packets into time bins and calculating multiple traffic statistics. Subsequently, we leverage Machine Learning (ML) technologies, training the algorithm on one dataset while testing the remaining two, to classify the loss presence in a bin. Despite the inherent hurdles posed by class imbalance and intricate traffic dynamics, we achieve decent outcomes (0.64 F1-score), effectively identifying the majority of lossy bins (0.64 recall) while guaranteeing the performance for lossless scenarios (0.94 recall).