TINYML NETWORK PROFILER IN BROWSER

Abstract

This research presents a client-side TinyML profiler of network in the web browser, designed for operation on resource-constrained devices and under unstable or low-bandwidth network conditions. The solution combines a threshold-based decision rule with a lightweight logistic (softmax) model executed locally to classify the quality of the network connection. Relevance. The rising share of mobile traffic, the heterogeneity of network environments, and the limited fidelity of existing browser indicators complicate accurate client-side decision-making. Object of study: client-side methods for profiling and classifying the quality of a browser-based network connection using standard Web Performance APIs and ML models. Aim: to analyze, design, and implement a profiler capable of classifying connection types and producing interpretable features without server support. Methods. Navigation/Resource Timing and PerformanceObserver are used to collect raw signals. A 14-dimensional feature vector is formed (medians/quantiles of RTT and throughput, variability measures, a loss-likelihood heuristic, and protocol/Service Worker indicators). The proposed a threshold rule with hysteresis and decision confidence, together with a softmax model. Results. The developed profiler requires no special permissions or third-party services. In controlled network-emulation scenarios it improves classification accuracy over a pure threshold baseline, while maintaining low overhead and decision explainability. Conclusions. The proposed client-side approach provides an effective, rapid, and interpretable assessment of network conditions in the browser and is suitable for resource-constrained settings. The results can be leveraged to further enhance adaptive content loading, expand the feature space, and deploy more compact ML models in web applications.