Semantic In-Band Network Telemetry for Low Bandwidth Overhead and High-Accuracy Monitoring

In-Band network telemetry (INT) offers fine-grained visibility by embedding per-packet monitoring data, which is essential for detecting anomalies in future 6G and cloud-native networks. However, naively collecting many telemetry items can cause severe traffic overhead, risking congestion and reduced network performance. To mitigate this, we propose semantic in-band network telemetry (SINT). SINT utilizes an autoencoder whose loss function is weighted according to each telemetry item's importance to anomaly detection. This enables selective compression of critical telemetry data at the data plane, significantly reducing traffic overhead while preserving detection accuracy. Experimental results on real-world mobile network traces show that SINT achieves an 87.5% reduction in traffic overhead with a 4% drop in accuracy.