acceptodds
Under review as a conference paper at ICLR 2027

CTU-Fusion: Utility-Guided Flow–Topology Fusion for Encrypted Traffic Classification

Abstract

Strong encrypted traffic classifiers remain constrained by the information within an individual flow. Historical and relational context can provide complementary evidence, but inappropriate fusion can corrupt correct predictions. We propose CTU-Fusion, a utility-guided Flow–Topology framework that separates primary prediction from auxiliary residual correction, estimates the sample-specific benefit and harm of each correction, and selectively controls its influence. Across three traffic-classification settings, CTU-Fusion improves Macro-F1 over the strongest evaluated baseline by 1.12–4.26%. Mapping-aware audits distinguish useful IoT history, negligible CICIDS topology, and severe negative transfer from unreliable USTC timestamps. Component, routing, and robustness studies separate Flow capability, complementary information, and realized fusion gains. These results show that relational evidence should influence a prediction only when its estimated utility is sufficient.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Encrypted Traffic Classification, Flow-Level Representation Learning, Topology-Aware Learning, Utility Estimation, Selective Information Fusion, Negative Transfer, Robustness

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_n4YQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes CTU-Fusion, for encrypted network traffic classification. It keeps a fixed single-flow classifier as the anchor, lets context experts propose a correction to its scores, and trains a judge that predicts whether applying the correction will fix a wrong answer or break a right one. Corrections are applied only when the predicted benefit clearly outweighs the predicted harm.

Critical strengths

The paper’s strongest point is its framing: it recognizes that adding relational context to a flow classifier can help, do nothing, or actively hurt, and it builds a clean permutation audit that shows all three cases on real datasets.

Critical weaknesses

The demonstrated overall perforamnce gain is minimun, as shown in Table 2. When standard devisiations are presented, which should presented togehter with the gverage values, there may not be any gain stantistically speaking for several of the presented cases. Lacking of sunstabtal performance enhancement is my biggest concern of this paper.

Questions for authors

none

Reviewer_QkCAOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies how to use historical and relational context for encrypted traffic classification without allowing unreliable context to damage correct predictions. CTU-Fusion freezes a Flow classifier, uses context-only experts to propose logit corrections, and trains a Utility estimator to predict the benefit and harm of accepting each correction. The evaluation uses Flow fallback on CICIDS2017, Host history with Utility routing on IoT-23, and session-set context on ISCX. It includes classification comparisons, ablations, context-mapping permutations, routing analysis, and robustness tests. An auxiliary USTC experiment illustrates negative transfer from forced context fusion, and the paper also reports external evaluations on UNSW and CTU-13.

Critical strengths

The paper addresses negative transfer in context fusion, which is a practical concern. Separating the original prediction from the context correction, and counting wrong-to-correct and correct-to-wrong transitions, helps explain what fusion changes. The IoT mapping-permutation experiment and routing audit directly examine whether history is relevant to the target sample and whether accepting corrections produces more benefit than harm.

Critical weaknesses
  1. The test set is used for policy tuning. The authors select thresholds, coverage, and β on the policy-fit portion of the final test partition, then report scores on the full partition, including those samples. The main results therefore include performance on data used to tune the policy and do not provide a clean estimate of performance on new data. The separate policy-audit provides some evidence, but the paper does not report complete results on untouched test data across datasets and seeds.

  2. Experimental validation is insufficient. The base models or processing configurations differ across settings: IoT uses an ensemble, ISCX uses session-level aggregation, and some experiments do not clearly specify the Flow model. The paper neither fixes an appropriate base model across scenarios nor compares different base models within the same scenario, leaving it unclear whether the gains depend on a particular configuration. In addition, CICIDS ultimately uses Flow alone, and ISCX context-only already matches the full method. The more complete evaluation of historical residual correction and Utility is concentrated on IoT, with only the Host relation enabled. The current experiments do not establish the framework’s effectiveness across scenarios and base models.

Questions for authors
  1. Why are thresholds, coverage, and β selected using the policy-fit portion of the final test set, followed by reporting results on the full test set? What results are obtained on untouched test data when all policy choices are made using validation data or a separate calibration set?

  2. The base models and configurations differ across settings: IoT uses an ensemble of two neural models and XGBoost, ISCX uses session-level aggregation, and some experiments do not clearly specify the Flow model. With the same appropriate Flow model held fixed, does CTU-Fusion improve performance across scenarios? Within the same scenario, do the gains persist when the base model is changed? Please use these comparisons to show whether the framework’s gains depend on a particular base model.

  3. Please define the historical statistics and all eight reliability features. What observation window does each of the payload, packet-sequence, and burst branches use? Are the 20 bursts extracted from the first 32 packets? Are all these features available when the prediction is made?

  4. What are the Utility model’s architecture, training losses, and reasons for choosing these inputs? Have you evaluated input-group ablations and the discrimination and calibration of the benefit/harm predictions? What role does the predicted gain play, and why is the harm weight set to 2?

  5. IoT uses an ensemble of two neural models and XGBoost as the Flow anchor, but Table 3 reports 79.29 for both Flow-only and the three-modality model. Which model does each result correspond to? Do all component ablations use the same frozen anchor?

  6. Please provide the missing USTC and external-evaluation settings. How are the USTC timestamps reconstructed, what establishes that timestamp distortion causes the performance drop, and how does Utility perform under the same conditions? For UNSW and CTU-13, what are the source training domains, label mappings, and scenario splits?

Reviewer_nhFvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

To mitigate negative transfer from naive topology fusion in encrypted traffic classification, this paper proposes CTU‑Fusion, which uses a frozen flow predictor, context‑derived residual corrections and an out‑of‑fold utility estimator to selectively apply topological adjustments with flow prediction fallback. Evaluated on public traffic datasets, it achieves 1.12‑4.26% Macro‑F1 gains with mapping‑aware audits and ablations, yet policy hyper‑parameters are calibrated on test‑set subsets and real‑world distribution‑shift validation remains limited.

Critical strengths

1: It identifies the under‑recognized negative‑transfer risk of naive topology fusion for encrypted traffic classification. Most context‑aware traffic models assume relational context is always helpful. This work separates flow capability, contextual signal and fusion performance, and proposes mapping‑aware audits to diagnose positive, negligible and harmful context regimes. This offers important empirical lessons for the traffic classification community.

2: Utility‑guided residual fusion features a practical no‑abstain fallback design. Unlike learn‑to‑defer methods which may skip predictions, CTU‑Fusion only rejects topology logit corrections and always outputs flow‑based predictions, fitting real‑time network monitoring requirements. Its out‑of‑fold pipeline enables sample‑wise estimation of correction benefit or harm.

Critical weaknesses

1: Flawed evaluation protocol weakens trust in core fusion gains. Hyper‑parameters for routing (threshold, β, coverage) are calibrated on a subset of the test partition, not a standalone validation set. Though a disjoint policy‑audit split exists, the primary reported main results include policy‑fit samples. We cannot verify whether utility‑routing improvements hold under strict fully‑held‑out evaluation, which is central to the paper’s main claim.

2: Generalization and safety of the utility estimator are insufficiently validated. The utility module needs sufficient beneficial/harmful transition examples for training and may struggle on rare classes. The Wilson‑UCB calibration is purely empirical with no formal risk guarantees. Limited public‑dataset domain‑shift experiments cannot convincingly demonstrate robustness against severe real‑world distribution shift.

Questions for authors
  1. Please report main comparison metrics only on the disjoint policy‑audit holdout split, with all hyper‑parameters tuned solely on policy‑fit data, to verify whether fusion gains persist under strict hold‑out evaluation.
  2. How does utility estimator performance degrade for rare classes with few benefit/harm transition samples? Provide per‑class prediction precision/recall for b^i,k\hat b_{i,k} and h^i,k\hat h_{i,k}.
  3. In the UNSW‑NB15 / CTU‑13 domain‑shift experiments: are acceptance thresholds reused from source domain or re‑calibrated on target data? Report the observed correct‑to‑wrong harmful correction rate on shifted data.
  4. Given ISCX only has 23 test sessions, quantify seed‑based variance for the reported 1.12% Macro‑F1 improvement, to rule out that the gain comes from random fluctuation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.