Incremental Learning in Network Traffic Management: A Fixed-Representation Rehearsal Approach
Francisco Rau, Petia Georgieva, Carlos Herranz, Iñaki Val, Joaquin Perez
MaxLinear, Inc., Valencia, Spain ·
Instituto de Telecomunicações / IEETA, University of Aveiro, Portugal ·
Electronic Engineering Department, Universitat de València, Spain
Keywords: Incremental Learning · Network Traffic Management · Deep Learning
Network traffic management systems must continuously evolve to new applications without degrading the performance of previously learned categories. This paper evaluates a class-incremental learning strategy that combines rehearsal-based replay with fixed feature representation. Starting from a Fully Connected Neural Network (FCNN) trained on six traffic classes using flow-level features, we freeze all feature-extraction layers and retrain only a newly initialized classification head. Incremental updates are performed using all samples from the incoming class(es) plus a compact, stratified rehearsal buffer containing 10% of the original training set. Experiments on 46,729 labeled flows assess two scenarios: adding one class and adding two classes, and compare against full retraining as an upper bound. The proposed approach reaches 87.90% and 87.72% test accuracy (weighted F1 of 0.877 and 0.875) with minimal forgetting (gaps up to 1.13%). Incremental updates reduce training time by 89–92% and decrease training-data storage requirements by up to 76.88%, supporting deployment in dynamic network environments.
29-dim flow features → pyramidal FCNN backbone (2048→1024→768→512→256, frozen after base training) → classification head retrained on rehearsal buffer + new-class samples.
⚡ New classes arrive: File Transfer + C2
Backbone (layers 1–5) stays frozen · 2 new output neurons initialized · head retrains on the rehearsal buffer (M) + new-class samples (scenario 6+2, 8 classes)
Forward pass (known class) Forward pass (new class) Rehearsal sample (10% replay) Frozen backbone weights (❄) Trainable head weights
Results — from scratch vs incremental
Training time
Time to train the model
from scratch → incremental : reduction
7 → 6 + 1 classes : 1,696 s → 133 s : 92.16%
8 → 6 + 2 classes : 1,872 s → 192 s : 89.72%
Storage
Memory of the training data
from scratch → incremental : reduction
7 → 6 + 1 classes : 9.2 MB → 2.1 MB : 76.88%
8 → 6 + 2 classes : 10.6 MB → 3.4 MB : 67.50%
Forgetting
Accuracy lost on old classes when adding new
old-class accuracy, from scratch → incremental : gap
7 → 6 + 1 classes : 86.4% → 85.8% : 0.56%
8 → 6 + 2 classes : 86.0% → 84.9% : 1.13%
As presented in the Lightning Talk (CSNDSP 2026). The live diagram above animates scenario 6+2 (File Transfer + C2 added together, 8 classes total).
Layer widths mirror the actual FCNN from the paper —
Input(29) → H1(2048) → H2(1024) → H3(768) → H4(512) → H5(256, output of fθ) → Head gφ(C classes, softmax).
Node counts per layer are downsampled for legibility; real neuron counts are labeled beneath each layer.
Periodically, two new classes ("File Transfer", "C2") are introduced — scenario 6+2 from the paper: only the head is
re-instantiated and retrained on Dinc = M ∪ Dnew (green pulses = samples drawn from the rehearsal buffer M,
shown bottom-right), while θ (backbone) remains frozen — exactly as in Eq. (1)–(2) of the paper.