Adaptive Semantic Compression for Programmable Data Plane Telemetry

Building on Sahin et al.'s use of In-band Network Telemetry (INT) for DDoS detection and the packet-level feature extraction in P4DDLe (Corin et al.), this idea proposes a programmable data plane framework that adapts the granularity and compression of telemetry data based on live anomaly scores. Unlike static compression (which risks information loss during critical events), the system would use lightweight, in-switch ML models (see Zhang et al.'s quantized inference toolbox) to assess risk and automatically increase semantic richness (e.g., more packet-level features, longer flow context) when anomalies are suspected. This enables richer data for downstream analytics only when needed, balancing overhead and detection accuracy. The novelty lies in real-time, data-driven adaptation of telemetry, which could enhance the efficacy and scalability of anomaly detection frameworks beyond what is currently possible with fixed telemetry pipelines.

References:

1. Leveraging In-band Network Telemetry for Automated DDoS Detection in Production Programmable Networks: The AmLight Use Case. Hadi Sahin, Jeronimo Bezerra, I. Brito, Renata Frez, Vasilka Chergarova, Luis Fernandez Lopez, J. Ibarra (2024). SC24-W: Workshops of the International Conference for High Performance Computing, Networking, Storage and Analysis.
2. Introducing Packet-Level Analysis in Programmable Data Planes to Advance Network Intrusion Detection. R. D. Corin, Luis Augusto Dias Knob, Luca Mendozzi, D. Siracusa, Marco Savi (2023). Comput. Networks.
3. A Machine Learning-Based Toolbox for P4 Programmable Data-Planes. Kaiyi Zhang, N. Samaan, Ahmed Karmouch (2024). IEEE Transactions on Network and Service Management.