Resolving Packets from Counters: Enabling Multi-scale Network Traffic Super Resolution...
summary
ZOOMSYNTH reconstructs packet-level traces from second‑level counters by stacking a transformer that models trends with an extreme‑value branch into a tree of upscalers, enabling multi‑stage super‑resolution. The system uses GTT modules assembled into a CLTM tree, adds a CLIP‑style rule adaptor for ACLs, and releases a 1.8B pretrained model trained on 16 GB (expanded to 1.7 TB counters). Results show quality comparable to packet-input baselines, real‑time synthesis at up to 1e9× upscaling within one second, +27.5% anomaly detection and +9.8% service recognition gains.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 2:05Fine-grained traffic traces challenges
- 3:31Traffic vs image super resolution
- 4:39ZoomSynth framework
- 6:07Granular Traffic Transformer
- 7:55Composable Large Traffic Model
- 9:00Low-rank fine-tuning
- 10:46Evaluation methodology
- 12:13Real-time performance
- 13:38Controlled generation results
- 14:41Conclusion
related talks
16mEnhancing Global Network Monitoring with Magnifier
15mCascadeNet: Generating Network Traffic with High-Fidelity Temporal Patterns
17mSupercharging Packet-level Network Simulation of Large Model Training via Memoization...
24mSuper Resolution: The scaler of tomorrow, here today!
24mConfluo: Distributed Monitoring and Diagnosis Stack for High-speed Networks
17mYuZu: Neural-Enhanced Volumetric Video Streaming
25mStroboscope: Declarative Network Monitoring on a Budget
17mDefeating Slow-and-Low Threats via Diffusion Model-based Generative Inference