Answers are grounded in this paper and cite their sources.
Smart building networks use heterogeneous Customer Premises Equipment (CPE), where telemetry is conditioned by manufacturers and operators, limiting network observability. This work presents an adaptive edge-based network telemetry framework combining external probes with native CPE telemetry and contextual telemetry, dynamically adjusting collection frequency based on network status and edge resources. The framework was evaluated using six hundred and thirty controlled physical injections across CPEs with different observability levels and seven event classes. SNMP and eBPF were used for observable CPE. Six of the seven classes showed recall values between 96.7% and 100%, while NET_SATURATION reached 86.7% to 100% using a dedicated iPerf3 server. Contextual telemetry allowed the detection of cross-activations between related events. The adaptive controller exhibited consistent patterns across different environments, and native telemetry allowed for the addition of complementary information for specific events. Compared with fixed high-intensity sampling, adaptive sampling maintained similar detection performance while reducing continuous probing, at the cost of longer confirmation times for DNS and HTTP events. Component ablation showed that persistence was essential for preserving detection precision and F1-score. The framework consumed 0.9–6.0% of one CPU core and 435–533 MB of memory, supporting the feasibility of edge-based telemetry in heterogeneous networks with limited observability.
No takes yet. Share an insight, caveat, or question.
Duarte et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: