acceptodds
Under review as a conference paper at ICLR 2027

PACC: Protocol-Aware Cross-Layer Compression for Compact Network Traffic Representation

Abstract

Network traffic classification underpins network security and management but faces growing challenges from pervasive encryption and evolving protocols. Existing representations involve clear trade-offs: hand-crafted flow statistics are efficient but lossy, raw-bit encodings can be accurate but costly, and pre-trained embeddings support transfer but often flatten the protocol stack and entangle cross-layer signals. We observe substantial redundancy both across and within network layers, which existing paradigms do not explicitly address, wasting capacity and encouraging shortcut learning that degrades generalization. To address this, we propose PACC, a redundancy-aware cross-layer representation framework. PACC models protocol layers as multiview inputs and learns faithful, compact projections factorized into shared (cross-layer) and private (layer-specific) components. It incorporates an information-preserving sparse segment representation to reduce structural repetition and supervised intra-layer filtering to suppress task-irrelevant semantic redundancy. A behavior-guided shared/private decomposition aligns cross-layer consensus while retaining layer-specific information, supported by information-theoretic analysis. Across encrypted application classification, IoT device identification, benign/malicious traffic classification, and Tor traffic detection, PACC achieves the best overall performance. On encrypted traffic, it achieves an F1 score of \(0.9685\), substantially outperforming existing baselines while using approximately \(15\times\) fewer computes and less GPU memory than the latest pre-trained traffic model.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Network Traffic Analysis, Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GLoUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes PACC that learn representation for encrypted network traffic. Each protocol layer in the networking stack is treated as a separate view. PACC explores the redundancy withing a layer and exploit consistency across layers. Four datasets are used to evaluate the performance and the application classification shows substantial enhancement while the other three tasks show minimum enhancement if any by considering statictial signifiance.

Critical strengths

The vision of removing redundancy and exploiting consistency within and among netowrking protocols at different layers is a great idea that has potential to bring great benefits.

The performance enhance on the applicaiton classification is substantial.

Critical weaknesses

This pape is motivated by protocol redundancy and cross-layer consistency, but never reutrns its findings back to networking knowledge.

Performance evaluaton in table 1 lacks standard deviation when the average is presented. Statistical significnance is not established, especially when comparng performance on tasks 2-4.

Investigate how much of the performace gain on task 1 comes from the exploititation of the proposed principle of removing redundance and exploiting consistency. The baselines' perforamnce is too bad to belive when compared with MLP. Make sure that the baseline is correctly and fairly configured.

Questions for authors

none

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

This paper proposes PACC, a protocol-aware cross-layer representation learning framework for network traffic classification that aims to reduce structural repetition and task-irrelevant information within protocol layers while preserving cross-layer shared information and layer-specific discriminative cues. PACC uses run-length encoding to compress repetitive content in protocol headers and combines structural reconstruction, classification supervision, and mutual information constraints to learn compact intra-layer representations. It then uses a global behavioral representation derived from statistics such as packet lengths, directions, and inter-arrival times as an anchor to guide the learning of shared and private representations for each protocol layer, followed by adaptive fusion based on each branch’s predictions. The authors evaluate PACC on four tasks: encrypted application classification, IoT device identification, benign and malicious traffic classification, and Tor traffic detection. They also conduct component ablation studies, computational efficiency evaluations, and experiments on few-shot recognition and temporal generalization.

Critical strengths
  1. Design tailored to the task. The method uses protocol layering and global flow behavior to learn shared information across layers and discriminative cues specific to each layer. Ablation studies support the contribution of each component.

  2. Clear gains in computational efficiency. The method requires approximately 1/14.5 of NetMamba’s inference computation, demonstrating the value of compact representations.

Critical weaknesses
  1. The incremental contribution over FactorCL is not sufficiently substantiated.
From the references, FactorCL already introduces a definition of “multiview non-redundancy” and learns task-relevant shared and view-specific representations via mutual-information objectives. Meanwhile, the PACC’s Definition 2.1 and FactorCL’s core Definition 2 are mathematically the same, and the overall approach follows the same basic idea. Therefore, PACC’s methodological novelty must be justified by concrete change. The paper lacks a direct comparison with the closest decomposition-learning baseline.

  2. Whether shared and private representations learn different information is not adequately verified.
The paper intends the shared branch to capture cross-layer common information and the private branches to capture layer-specific classification cues. However, the current training objectives still allow both branches to retain the same information. Even if the two branches are fully redundant, the decomposition conditions stated in the paper can still be satisfied, and metrics meant to measure information leakage may still look favorable. Thus, these conditions and metrics are insufficient to confirm that the model achieves the intended information factorization. Ablations show that the branches and objectives help classification, but they do not cleanly separate gains from information decomposition, extra branches, and auxiliary supervision. Further analysis is needed on what each branch retains to verify that improvements come from the proposed decomposition mechanism rather than capacity or extra losses alone.

Questions for authors
  1. The core idea appears to build on FactorCL; some innovations look like adaptations of network-traffic properties into that framework or mindset. Could the authors clarify the necessity of these changes and for incremental benefit over FactorCL?

  2. Beyond the cross-date UNSW-IoT experiment and the within-split DoH evaluation, can the authors further characterize generalization to changes in the collection environment and the gap between training and independently held-out testing? On these public datasets, deep learning methods already perform very well; robustness to environment shift may be more informative for judging real-world effectiveness.

  3. The latency and throughput experiment exclude preprocessing, yet the reported preprocessing time is roughly 200× the inference latency, and baseline preprocessing costs are unreported. After including protocol parsing, feature construction, and compression for all methods, how much end-to-end latency and throughput advantage does PACC retain under the same resources?

  4. Do the shared and private branches learn different information? The current conditions and leakage metrics still permit redundant encodings. Can the authors run controlled experiments to inspect what each branch retains and compare against a conventional two-branch model with matched parameters, to confirm that gains come from the intended information decomposition?

Reviewer_q5ynOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PACC, a protocol‑aware cross‑layer representation framework for encrypted traffic classification. Treating protocol layers as multi‑view inputs, PACC reduces intra‑layer redundancy and decomposes features into shared cross‑layer consensus and layer‑specific private evidence anchored by flow‑level behavior, before adaptive fusion for classification. Evaluated on four traffic benchmarks, PACC achieves strong accuracy with much lower computational overhead, validated via ablations and generalization tests.

Critical strengths
  1. Well‑grounded information‑theoretic formulation for cross‑layer traffic representation. The authors formalize network protocol layers as multi‑view inputs and explicitly distinguish intra‑layer structural/semantic redundancy from inter‑layer cross‑layer consensus. Mutual‑information‑driven definitions, propositions and proofs provide solid theoretical underpinnings for the shared‑private decomposition design. This brings principled thinking to encrypted traffic representation learning, which is valuable for the research community.
  2. Strong empirical performance together with high computational efficiency. On the challenging CipherSpectrum encrypted application classification benchmark, PACC obtains substantial F1 improvement over strong baselines meanwhile reducing multiply‑accumulate operations by ~15×. Comprehensive ablation, hyper‑parameter sensitivity and generalization (few‑shot, temporal shift) experiments are conducted to validate module contributions.
Critical weaknesses
  1. Limited practical novelty in core architectural components despite solid theoretical analysis. The key building blocks including run‑length‑based lossless compression, information‑based denoising via CLUB estimator, InfoNCE‑driven shared‑private disentanglement and sample‑wise adaptive weighted fusion are well‑established techniques originated from existing multi‑view/multimodal learning literature. The main novelty lies in applying these existing tools to network traffic cross‑layer setting, rather than introducing substantially new neural mechanisms or learning objectives tailored for traffic data. The paper fails to sufficiently highlight what new algorithmic insight is uniquely invented for protocol‑aware traffic modeling.
  2. Real‑world deployment and adversarial robustness evaluation is insufficient. Modern encrypted traffic classification systems face realistic threats including traffic obfuscation, packet mangling and concept drift induced by protocol updates. While the paper reports temporal generalization and few‑shot results, there is no evaluation against adversarial manipulations or real‑world noisy packet captures beyond curated public datasets. Without such evidence, it remains unclear whether the promising accuracy‑efficiency trade‑off observed on benchmark datasets can transfer to real operational environments, which significantly undermines the practical impact of this traffic‑security‑oriented work.
Questions for authors
  1. You adopt multiple existing components (run‑length compression, CLUB penalty, InfoNCE shared‑private decomposition) from prior multimodal learning works. Could you explicitly articulate what new algorithmic modifications or insights are uniquely proposed for network protocol traffic, beyond re‑purposing existing multi‑view techniques? Please quantitatively contrast your design against a strong multi‑view baseline adapted for traffic to demonstrate your unique added value.
  2. Your experiments are mainly conducted on cleaned public datasets. How would PACC behave when input packets contain real‑world artifacts such as corrupted headers, partial packet fragments, or traffic obfuscation that perturbs header fields and packet‑level behavioral statistics? Could you provide concrete experimental results under these perturbation settings?
  3. The flow‑behavior anchor ZgZ_g plays a central role to guide shared‑private decomposition. What will happen if behavioral side‑channels are heavily distorted (e.g., traffic padding to obfuscate packet sizes)? How fragile is the whole decomposition pipeline to anchor quality?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.