GESR: Graph-Based Edge Semantic Reconstruction for Stealthy Communication Detection with Benign-Only Training

Henghui Xu; Xiaobo Ma; Yuchen Zhang

arxiv: 2605.07536 · v1 · submitted 2026-05-08 · 💻 cs.CR · cs.LG

GESR: Graph-Based Edge Semantic Reconstruction for Stealthy Communication Detection with Benign-Only Training

Henghui Xu , Yuchen Zhang , Xiaobo Ma This is my paper

Pith reviewed 2026-05-11 01:57 UTC · model grok-4.3

classification 💻 cs.CR cs.LG

keywords graph anomaly detectionnetwork intrusion detectionbenign-only trainingedge semantic reconstructionstealthy communicationstructural consistencyhost-level scoringCICIDS2017

0 comments

The pith

Reconstructing expected edge semantics from local neighborhood topology detects stealthy malicious communications with only benign training data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper seeks to identify stealthy malicious network flows that mimic normal traffic like HTTPS without using any attack examples in training. Conventional detectors either demand labeled malicious data or score each flow in isolation, missing the relational context that reveals suspicious activity. GESR builds attributed communication graphs and reconstructs what each edge should semantically represent based purely on its surrounding structural neighborhood. Inconsistencies between predicted and observed edges are then aggregated into host-level anomaly scores through median absolute deviation scaling. This approach matters because it targets context-dependent threats that evade independent feature analysis while raising the bar for attackers who must now also match expected neighborhood patterns.

Core claim

GESR models complex network activity as attributed communication graphs and reconstructs edge semantics entirely from local structural context rather than isolated features. This design forces prediction of expected communication patterns from neighborhood topologies, which attackers cannot easily manipulate. The resulting structural inconsistencies are converted into host-level anomaly scores using robust Median Absolute Deviation calibration, enabling detection of suspicious communications and anomalous hosts under a benign-only training setting.

What carries the argument

Edge semantic reconstruction module that predicts expected communication patterns solely from local neighborhood topologies in attributed communication graphs.

If this is right

Detects sparse, context-dependent suspicious activity without any labeled attack examples.
Raises the difficulty of evasion because attackers must satisfy neighborhood topology expectations in addition to flow features.
Converts per-edge inconsistencies into calibrated host-level anomaly scores via median absolute deviation.
Achieves ROC-AUC of 0.9753 and true-positive rate of 0.8569 at 5% false-positive rate on CICIDS2017.
Outperforms prior methods on both CTU-13 and CICIDS2017 under tight false-positive constraints.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same local-structure reconstruction principle could transfer to anomaly detection in other graph domains such as transaction or sensor networks.
Dynamic graphs with time-varying neighborhoods would require extensions to maintain reconstruction accuracy over long periods.
Hybrid systems that combine edge reconstruction with lightweight feature checks might further lower false positives in production networks.
Testing on larger, real-world enterprise traffic logs would reveal whether the structural dependency scales beyond the evaluated benchmarks.

Load-bearing premise

That structural inconsistencies between reconstructed and actual edges correspond to malicious activity rather than unusual but legitimate communication patterns.

What would settle it

Malicious flows that produce reconstructed edge semantics matching their actual neighborhood topology so closely that they receive low anomaly scores at the 5% FPR operating point.

Figures

Figures reproduced from arXiv: 2605.07536 by Henghui Xu, Xiaobo Ma, Yuchen Zhang.

**Figure 1.** Figure 1: GESR system overview. During training, GESR builds attributed communication graphs from benign flow records and learns structure-conditioned edge semantic reconstruction with an edge-aware encoder. During inference, the same encoder scores anomalous communications, calibrates edge scores, and aggregates them to produce ranked suspicious hosts. This deployment perspective matters because intrusion datasets… view at source ↗

read the original abstract

Detecting stealthy malicious communications from flow logs under benign-only training remains a critical challenge in network security. Malicious communications often camouflage as normal traffic like standard HTTPS flows. Conventional intrusion detectors rely strictly on known labeled attacks. Alternatively, they score flows completely independently. These approaches fail against sparse and context-dependent suspicious activity. To capture this essential context, graph anomaly detectors have been introduced to add valuable relational information to the analysis. However, existing methods fail to test the structural consistency of specific communication edges. To overcome these fundamental limitations, we present GESR, a novel graph-based framework for detecting suspicious communications and anomalous hosts under a benign-only training setting. GESR models complex network activity as attributed communication graphs. It cleverly reconstructs edge semantics entirely from local structural context rather than isolated features. This non-intuitive design forces the framework to predict expected communication patterns from neighborhood topologies. Attackers cannot easily manipulate this deep structural dependency. The model then converts the resulting structural inconsistencies into host-level anomaly scores. It utilizes robust Median Absolute Deviation (MAD) calibration for this final step. We evaluate GESR extensively on CTU-13 and CICIDS2017 datasets. These evaluations strictly impose tight false-positive operating constraints. On CICIDS2017, GESR achieves an outstanding ROC-AUC of 0.9753. It also yields a high TPR of 0.8569 at a strict 5% FPR threshold. GESR consistently outperforms existing methods across both evaluated benchmarks. The results prove that structure-conditioned edge reconstruction is a credible direction for practical intrusion detection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

GESR adds structure-conditioned edge reconstruction to graph anomaly detection for benign-only network intrusion detection and reports competitive numbers on standard datasets, but the core claim that reconstruction errors reliably signal malice lacks supporting analysis.

read the letter

The main thing to know is that GESR models network flows as attributed graphs and trains a model to reconstruct edge semantics purely from local neighborhood topology under benign-only data. It then converts reconstruction errors into host-level scores using MAD calibration. On CICIDS2017 it reaches 0.975 ROC-AUC and 0.857 TPR at 5% FPR while beating the baselines they compare against; similar gains appear on CTU-13 under tight FPR constraints. That is the concrete result worth noting. The approach is new in its insistence that the reconstruction must come from structural context rather than isolated features, which in principle makes it harder for an attacker to blend in without matching neighborhood expectations. The focus on low false-positive regimes also matches real deployment needs. The paper does a reasonable job framing the problem of context-dependent stealthy attacks and giving a workable pipeline. The soft spots are more substantial. No architecture details, graph construction rules, loss functions, or ablation results are visible, so it is impossible to judge whether the performance depends on careful tuning or holds more generally. The central assumption—that structural inconsistencies detected this way correspond to malice rather than unusual but legitimate traffic patterns—receives no direct test. There is also no adversarial evaluation showing whether an attacker who observes common neighborhood structures can keep reconstruction error low. The MAD step fitted on the same benign data introduces a minor risk of optimistic scoring. This paper is aimed at researchers working on graph-based or unsupervised intrusion detection. A reader already active in that niche could extract the reconstruction idea and try it, but anyone outside the subfield or looking for immediately deployable methods would get limited value. I would send it to peer review. The numbers are specific enough and the framing coherent enough that referees should see the full methods and experiments to pressure-test the assumptions.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces GESR, a graph-based framework for detecting stealthy malicious communications in network flow logs under benign-only training. Network activity is modeled as attributed communication graphs; edge semantics are reconstructed solely from local neighborhood topologies rather than isolated features. Structural inconsistencies are aggregated into host-level anomaly scores using Median Absolute Deviation (MAD) calibration. The method is evaluated on CTU-13 and CICIDS2017 under tight FPR constraints, reporting ROC-AUC 0.9753 and TPR 0.8569 at 5% FPR on CICIDS2017 while outperforming baselines.

Significance. If the central claim holds, GESR would represent a meaningful advance in unsupervised network anomaly detection by exploiting relational structure to identify context-dependent anomalies that evade per-flow feature detectors. The benign-only training regime and emphasis on neighborhood-conditioned reconstruction address practical constraints in real deployments. The reported performance under strict FPR limits and the use of robust MAD calibration are positive elements that could support practical utility if the underlying assumptions are validated.

major comments (3)

[§3] §3 (Method): The core claim that edge reconstruction from local structural context forces attackers to satisfy neighborhood topology expectations is load-bearing, yet the manuscript provides no explicit description of the reconstruction model architecture, loss function, or how the attributed graph is constructed from flow logs. Without these details it is impossible to verify whether the reported gains derive from the proposed structural dependency or from other implementation choices.
[§4] §4 (Evaluation): The CICIDS2017 results (ROC-AUC 0.9753, TPR 0.8569 at 5% FPR) are presented without ablation studies isolating the contribution of neighborhood-conditioned reconstruction versus simpler graph or feature baselines, nor any analysis of benign traffic diversity (e.g., varying host roles or application patterns that naturally alter local topologies). This leaves open whether reconstruction error reliably signals malice rather than benign structural variation.
[§3.3] §3.3 (Host Scoring): The final MAD calibration step is described as converting edge inconsistencies to host scores, but the manuscript does not clarify whether MAD parameters are estimated exclusively on held-out benign data or on the same training set used for the reconstruction model. If the latter, the scoring procedure risks circularity that could inflate the reported metrics.

minor comments (2)

[Abstract] The abstract and introduction would benefit from a concise statement of the exact graph construction procedure (node/edge attributes, aggregation windows) to allow readers to reproduce the experimental setup.
[§4] Figure captions and table footnotes should explicitly state the number of runs, random seeds, and confidence intervals for all reported metrics to improve reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. We address each major point below and will incorporate revisions to improve clarity, reproducibility, and rigor.

read point-by-point responses

Referee: [§3] §3 (Method): The core claim that edge reconstruction from local structural context forces attackers to satisfy neighborhood topology expectations is load-bearing, yet the manuscript provides no explicit description of the reconstruction model architecture, loss function, or how the attributed graph is constructed from flow logs. Without these details it is impossible to verify whether the reported gains derive from the proposed structural dependency or from other implementation choices.

Authors: We agree that Section 3 currently provides only a high-level description of the GESR framework and does not include the specific architectural details, loss function, or graph construction procedure needed for full verification and reproducibility. In the revised manuscript we will expand Section 3 with: (1) the precise steps for building the attributed communication graph from raw flow logs (node/edge attributes and aggregation rules); (2) the reconstruction model architecture (graph neural network layers, dimensions, and activation functions); and (3) the training loss function (including its mathematical formulation). These additions will directly substantiate the structural-dependency claim. revision: yes
Referee: [§4] §4 (Evaluation): The CICIDS2017 results (ROC-AUC 0.9753, TPR 0.8569 at 5% FPR) are presented without ablation studies isolating the contribution of neighborhood-conditioned reconstruction versus simpler graph or feature baselines, nor any analysis of benign traffic diversity (e.g., varying host roles or application patterns that naturally alter local topologies). This leaves open whether reconstruction error reliably signals malice rather than benign structural variation.

Authors: We concur that the current evaluation lacks the ablations and benign-diversity analysis required to isolate the benefit of neighborhood-conditioned reconstruction. The revised Section 4 will add: (1) ablation experiments comparing full GESR against feature-only and non-graph baselines; (2) quantitative analysis of reconstruction-error distributions across diverse benign host roles and application patterns in CICIDS2017; and (3) statistical comparison showing that malicious edges produce significantly higher errors than the observed benign topological variations. This will strengthen the claim that the reported performance stems from the proposed structural mechanism. revision: yes
Referee: [§3.3] §3.3 (Host Scoring): The final MAD calibration step is described as converting edge inconsistencies to host scores, but the manuscript does not clarify whether MAD parameters are estimated exclusively on held-out benign data or on the same training set used for the reconstruction model. If the latter, the scoring procedure risks circularity that could inflate the reported metrics.

Authors: We appreciate the referee highlighting this ambiguity. The manuscript does not currently specify the data used for MAD estimation. In the revision we will explicitly state in Section 3.3 that MAD parameters are computed exclusively on a held-out subset of benign data that is disjoint from the training set used for the reconstruction model. This clarification removes any possibility of circularity and ensures the host-level scores are derived from an independent calibration. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper trains a graph model on benign-only data to reconstruct edge semantics from neighborhood topologies, then applies standard MAD calibration to the resulting reconstruction errors to produce host-level anomaly scores. This is a conventional unsupervised anomaly detection pipeline: the reconstruction acts as a learned normalcy model, and deviations are scored externally on labeled test sets (CICIDS2017, CTU-13). No equations or steps reduce the claimed prediction or final score to the training inputs by construction, no self-citations bear the central load, and no ansatz or uniqueness claim is smuggled in. The reported ROC-AUC and TPR are computed against ground-truth labels on held-out data, keeping the evaluation independent of the fitted parameters.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unstated assumption that network communication graphs can be constructed from flow logs in a way that preserves structural signals of malice, plus the choice of MAD for final scoring. No free parameters or invented entities are explicitly named in the abstract.

axioms (1)

domain assumption Attributed communication graphs can be built from flow logs such that local topology encodes expected benign patterns.
Invoked when the paper states that the model reconstructs edge semantics from neighborhood topologies.

pith-pipeline@v0.9.0 · 5591 in / 1448 out tokens · 27442 ms · 2026-05-11T01:57:34.220418+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

GESR models complex network activity as attributed communication graphs. It cleverly reconstructs edge semantics entirely from local structural context rather than isolated features... uses an edge-aware message-passing encoder based on GINEConv... masked reconstruction objective... anomaly scores according to the discrepancy between observed edge semantics and their reconstructed values... MAD calibration
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

structure-conditioned edge semantic reconstruction... structural inconsistencies into host-level anomaly scores

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages

[1]

Xu et al

Parssegny, C., Mazel, J., Levillain, O., Chifflier, P.: Striking Back at Cobalt: Using Network Traffic Metadata to Detect Cobalt Strike Masquerading Command and 18 H. Xu et al. Control Channels. In: Dalla Preda, M., Schrittwieser, S., Naessens, V., De Sut- ter, B. (eds.) Availability, Reliability and Security. ARES 2025. Lecture Notes in Computer Science,...

work page 2025
[2]

Cybersecurity2, 20 (2019)

Khraisat, A., Gondal, I., Vamplew, P., Kamruzzaman, J.: Survey of Intrusion De- tection Systems: Techniques, Datasets and Challenges. Cybersecurity2, 20 (2019)

work page 2019
[3]

Buczak, A.L., Guven, E.: A Survey of Data Mining and Machine Learning Methods forCyberSecurityIntrusionDetection.IEEECommunicationsSurveys&Tutorials 18(2), 1153–1176 (2016)

work page 2016
[4]

In: IEEE Symposium on Security and Privacy (S&P), pp

Sommer, R., Paxson, V.: Outside the Closed World: On Using Machine Learning for Network Intrusion Detection. In: IEEE Symposium on Security and Privacy (S&P), pp. 305–316. IEEE (2010)

work page 2010
[5]

Computer Networks51(12), 3448–3470 (2007)

Patcha, A., Park, J.-M.: An Overview of Anomaly Detection Techniques: Existing Solutions and Latest Technological Trends. Computer Networks51(12), 3448–3470 (2007)

work page 2007
[6]

Internet Society (2018)

Mirsky, Y., Doitshman, T., Elovici, Y., Shabtai, A.: Kitsune: An Ensemble of AutoencodersforOnlineNetworkIntrusionDetection.In:NetworkandDistributed System Security Symposium (NDSS). Internet Society (2018)

work page 2018
[7]

Data Mining and Knowledge Discovery29(3), 626–688 (2015)

Akoglu, L., Tong, H., Koutra, D.: Graph-based Anomaly Detection and Descrip- tion: A Survey. Data Mining and Knowledge Discovery29(3), 626–688 (2015)

work page 2015
[8]

Computers & Security45, 100–123 (2014)

Garcia, S., Grill, M., Stiborek, J., Zunino, A.: An empirical comparison of botnet detection methods. Computers & Security45, 100–123 (2014)

work page 2014
[9]

In: ICISSP, pp

Sharafaldin, I., Lashkari, A.H., Ghorbani, A.A.: Toward Generating a New In- trusion Detection Dataset and Intrusion Traffic Characterization. In: ICISSP, pp. 108–116. SciTePress (2018)

work page 2018
[10]

Computers & Security86, 147–167 (2019)

Ring, M., Wunderlich, S., Scheuring, D., Landes, D., Hotho, A.: A Survey of Network-Based Intrusion Detection Data Sets. Computers & Security86, 147–167 (2019)

work page 2019
[11]

In: ICDM, pp

Liu, F.T., Ting, K.M., Zhou, Z.H.: Isolation Forest. In: ICDM, pp. 413–422. IEEE (2008)

work page 2008
[12]

In: NIPS, pp

Hamilton, W.L., Ying, Z., Leskovec, J.: Inductive Representation Learning on Large Graphs. In: NIPS, pp. 1024–1034 (2017)

work page 2017
[13]

In: ICLR (2020)

Hu, W., Liu, B., Gomes, J., Zitnik, J., Liang, P., Pande, V., Leskovec, J.: Strategies for Pre-training Graph Neural Networks. In: ICLR (2020)

work page 2020
[14]

In: Proceedings of the 2019 SIAM International Conference on Data Mining (SDM), pp

Ding, K., Li, J., Bhanushali, R., Liu, H.: Deep Anomaly Detection on Attributed Networks. In: Proceedings of the 2019 SIAM International Conference on Data Mining (SDM), pp. 594–602. SIAM (2019)

work page 2019
[15]

IEEE Transactions on Neural Networks and Learning Systems33(6), 2378–2392 (2022)

Liu, Y., Li, Z., Pan, S., Gong, C., Zhou, C., Karypis, G.: Anomaly Detection on Attributed Networks via Contrastive Self-Supervised Learning. IEEE Transactions on Neural Networks and Learning Systems33(6), 2378–2392 (2022)

work page 2022
[16]

In: Proceedings of the 30th ACM International Conference on Information & Knowledge Management (CIKM), pp

Jin,M.,Liu,Y.,Zheng,Y.,Chi,L.,Li,Y.-F.,Pan,S.:ANEMONE:GraphAnomaly Detection with Multi-Scale Contrastive Learning. In: Proceedings of the 30th ACM International Conference on Information & Knowledge Management (CIKM), pp. 3122–3126. ACM (2021)

work page 2021
[17]

In: Proceedings of the 2016 ACM Workshop on Artificial Intelli- gence and Security (AISec), pp

Anderson, B., McGrew, D.A.: Identifying Encrypted Malware Traffic with Contex- tual Flow Data. In: Proceedings of the 2016 ACM Workshop on Artificial Intelli- gence and Security (AISec), pp. 35–46. ACM (2016)

work page 2016
[18]

In: Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), pp

Anderson, B., McGrew, D.A.: Machine Learning for Encrypted Malware Traffic Classification: Accounting for Noisy Labels and Non-Stationarity. In: Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), pp. 1723–1732. ACM (2017) GESRfor Stealthy Communication Detection 19

work page 2017
[19]

Journal of Computer Virology and Hacking Techniques14(3), 195– 211 (2018)

Anderson, B., Paul, S., McGrew, D.: Deciphering Malware’s Use of TLS (Without Decryption). Journal of Computer Virology and Hacking Techniques14(3), 195– 211 (2018)

work page 2018
[20]

Knowledge-Based Systems258, 110030 (2022)

Caville, E., Lo, W.W., Layeghy, S., Portmann, M.: Anomal-E: A Self- Supervised Network Intrusion Detection System Based on Graph Neural Networks. Knowledge-Based Systems258, 110030 (2022)

work page 2022
[21]

PLoS ONE10(3), e0118432 (2015)

Saito, T., Rehmsmeier, M.: The Precision-Recall Plot Is More Informative than the ROC Plot When Evaluating Binary Classifiers on Imbalanced Datasets. PLoS ONE10(3), e0118432 (2015)

work page 2015
[22]

In: Proceedings of the Network and Distributed System Security Symposium (NDSS)

Gu, G., Zhang, J., Lee, W.: BotSniffer: Detecting Botnet Command and Con- trol Channels in Network Traffic. In: Proceedings of the Network and Distributed System Security Symposium (NDSS). Internet Society (2008)

work page 2008
[23]

In: Proceed- ings of the 17th USENIX Security Symposium, pp

Gu, G., Perdisci, R., Zhang, J., Lee, W.: BotMiner: Clustering Analysis of Network Traffic for Protocol- and Structure-Independent Botnet Detection. In: Proceed- ings of the 17th USENIX Security Symposium, pp. 139–154. USENIX Association (2008)

work page 2008
[24]

In: Proceedings of the 19th USENIX Security Symposium, pp

Nagaraja, S., Mittal, P., Hong, C.-Y., Caesar, M., Borisov, N.: BotGrep: Finding P2P Bots with Structured Graph Analysis. In: Proceedings of the 19th USENIX Security Symposium, pp. 95–110. USENIX Association (2010)

work page 2010
[25]

Information Sciences644, 119229 (2023)

Hong, Y., Li, Q., Yang, Y., Shen, M.: Graph based Encrypted Malicious Traffic Detection with Hybrid Analysis of Multi-view Features. Information Sciences644, 119229 (2023)

work page 2023
[26]

In: Proceedings of the 25th International Symposium on Research in Attacks, Intrusions and Defenses (RAID), pp

Fu, Z., Liu, M., Qin, Y., Zhang, J., Zou, Y., Yin, Q., Li, Q., Duan, H.: Encrypted Malware Traffic Detection via Graph-based Network Analysis. In: Proceedings of the 25th International Symposium on Research in Attacks, Intrusions and Defenses (RAID), pp. 495–509. ACM (2022)

work page 2022
[27]

In: Proceedings of the 26th International Symposium on Research in Attacks, Intrusions and Defenses (RAID), pp

King, I.J., Shu, X., Jang, J., Eykholt, K., Lee, T., Huang, H.H.: EdgeTorrent: Real- time Temporal Graph Representations for Intrusion Detection. In: Proceedings of the 26th International Symposium on Research in Attacks, Intrusions and Defenses (RAID), pp. 77–91. ACM (2023)

work page 2023

[1] [1]

Xu et al

Parssegny, C., Mazel, J., Levillain, O., Chifflier, P.: Striking Back at Cobalt: Using Network Traffic Metadata to Detect Cobalt Strike Masquerading Command and 18 H. Xu et al. Control Channels. In: Dalla Preda, M., Schrittwieser, S., Naessens, V., De Sut- ter, B. (eds.) Availability, Reliability and Security. ARES 2025. Lecture Notes in Computer Science,...

work page 2025

[2] [2]

Cybersecurity2, 20 (2019)

Khraisat, A., Gondal, I., Vamplew, P., Kamruzzaman, J.: Survey of Intrusion De- tection Systems: Techniques, Datasets and Challenges. Cybersecurity2, 20 (2019)

work page 2019

[3] [3]

Buczak, A.L., Guven, E.: A Survey of Data Mining and Machine Learning Methods forCyberSecurityIntrusionDetection.IEEECommunicationsSurveys&Tutorials 18(2), 1153–1176 (2016)

work page 2016

[4] [4]

In: IEEE Symposium on Security and Privacy (S&P), pp

Sommer, R., Paxson, V.: Outside the Closed World: On Using Machine Learning for Network Intrusion Detection. In: IEEE Symposium on Security and Privacy (S&P), pp. 305–316. IEEE (2010)

work page 2010

[5] [5]

Computer Networks51(12), 3448–3470 (2007)

Patcha, A., Park, J.-M.: An Overview of Anomaly Detection Techniques: Existing Solutions and Latest Technological Trends. Computer Networks51(12), 3448–3470 (2007)

work page 2007

[6] [6]

Internet Society (2018)

Mirsky, Y., Doitshman, T., Elovici, Y., Shabtai, A.: Kitsune: An Ensemble of AutoencodersforOnlineNetworkIntrusionDetection.In:NetworkandDistributed System Security Symposium (NDSS). Internet Society (2018)

work page 2018

[7] [7]

Data Mining and Knowledge Discovery29(3), 626–688 (2015)

Akoglu, L., Tong, H., Koutra, D.: Graph-based Anomaly Detection and Descrip- tion: A Survey. Data Mining and Knowledge Discovery29(3), 626–688 (2015)

work page 2015

[8] [8]

Computers & Security45, 100–123 (2014)

Garcia, S., Grill, M., Stiborek, J., Zunino, A.: An empirical comparison of botnet detection methods. Computers & Security45, 100–123 (2014)

work page 2014

[9] [9]

In: ICISSP, pp

Sharafaldin, I., Lashkari, A.H., Ghorbani, A.A.: Toward Generating a New In- trusion Detection Dataset and Intrusion Traffic Characterization. In: ICISSP, pp. 108–116. SciTePress (2018)

work page 2018

[10] [10]

Computers & Security86, 147–167 (2019)

Ring, M., Wunderlich, S., Scheuring, D., Landes, D., Hotho, A.: A Survey of Network-Based Intrusion Detection Data Sets. Computers & Security86, 147–167 (2019)

work page 2019

[11] [11]

In: ICDM, pp

Liu, F.T., Ting, K.M., Zhou, Z.H.: Isolation Forest. In: ICDM, pp. 413–422. IEEE (2008)

work page 2008

[12] [12]

In: NIPS, pp

Hamilton, W.L., Ying, Z., Leskovec, J.: Inductive Representation Learning on Large Graphs. In: NIPS, pp. 1024–1034 (2017)

work page 2017

[13] [13]

In: ICLR (2020)

Hu, W., Liu, B., Gomes, J., Zitnik, J., Liang, P., Pande, V., Leskovec, J.: Strategies for Pre-training Graph Neural Networks. In: ICLR (2020)

work page 2020

[14] [14]

In: Proceedings of the 2019 SIAM International Conference on Data Mining (SDM), pp

Ding, K., Li, J., Bhanushali, R., Liu, H.: Deep Anomaly Detection on Attributed Networks. In: Proceedings of the 2019 SIAM International Conference on Data Mining (SDM), pp. 594–602. SIAM (2019)

work page 2019

[15] [15]

IEEE Transactions on Neural Networks and Learning Systems33(6), 2378–2392 (2022)

Liu, Y., Li, Z., Pan, S., Gong, C., Zhou, C., Karypis, G.: Anomaly Detection on Attributed Networks via Contrastive Self-Supervised Learning. IEEE Transactions on Neural Networks and Learning Systems33(6), 2378–2392 (2022)

work page 2022

[16] [16]

In: Proceedings of the 30th ACM International Conference on Information & Knowledge Management (CIKM), pp

Jin,M.,Liu,Y.,Zheng,Y.,Chi,L.,Li,Y.-F.,Pan,S.:ANEMONE:GraphAnomaly Detection with Multi-Scale Contrastive Learning. In: Proceedings of the 30th ACM International Conference on Information & Knowledge Management (CIKM), pp. 3122–3126. ACM (2021)

work page 2021

[17] [17]

In: Proceedings of the 2016 ACM Workshop on Artificial Intelli- gence and Security (AISec), pp

Anderson, B., McGrew, D.A.: Identifying Encrypted Malware Traffic with Contex- tual Flow Data. In: Proceedings of the 2016 ACM Workshop on Artificial Intelli- gence and Security (AISec), pp. 35–46. ACM (2016)

work page 2016

[18] [18]

In: Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), pp

Anderson, B., McGrew, D.A.: Machine Learning for Encrypted Malware Traffic Classification: Accounting for Noisy Labels and Non-Stationarity. In: Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), pp. 1723–1732. ACM (2017) GESRfor Stealthy Communication Detection 19

work page 2017

[19] [19]

Journal of Computer Virology and Hacking Techniques14(3), 195– 211 (2018)

Anderson, B., Paul, S., McGrew, D.: Deciphering Malware’s Use of TLS (Without Decryption). Journal of Computer Virology and Hacking Techniques14(3), 195– 211 (2018)

work page 2018

[20] [20]

Knowledge-Based Systems258, 110030 (2022)

Caville, E., Lo, W.W., Layeghy, S., Portmann, M.: Anomal-E: A Self- Supervised Network Intrusion Detection System Based on Graph Neural Networks. Knowledge-Based Systems258, 110030 (2022)

work page 2022

[21] [21]

PLoS ONE10(3), e0118432 (2015)

Saito, T., Rehmsmeier, M.: The Precision-Recall Plot Is More Informative than the ROC Plot When Evaluating Binary Classifiers on Imbalanced Datasets. PLoS ONE10(3), e0118432 (2015)

work page 2015

[22] [22]

In: Proceedings of the Network and Distributed System Security Symposium (NDSS)

Gu, G., Zhang, J., Lee, W.: BotSniffer: Detecting Botnet Command and Con- trol Channels in Network Traffic. In: Proceedings of the Network and Distributed System Security Symposium (NDSS). Internet Society (2008)

work page 2008

[23] [23]

In: Proceed- ings of the 17th USENIX Security Symposium, pp

Gu, G., Perdisci, R., Zhang, J., Lee, W.: BotMiner: Clustering Analysis of Network Traffic for Protocol- and Structure-Independent Botnet Detection. In: Proceed- ings of the 17th USENIX Security Symposium, pp. 139–154. USENIX Association (2008)

work page 2008

[24] [24]

In: Proceedings of the 19th USENIX Security Symposium, pp

Nagaraja, S., Mittal, P., Hong, C.-Y., Caesar, M., Borisov, N.: BotGrep: Finding P2P Bots with Structured Graph Analysis. In: Proceedings of the 19th USENIX Security Symposium, pp. 95–110. USENIX Association (2010)

work page 2010

[25] [25]

Information Sciences644, 119229 (2023)

Hong, Y., Li, Q., Yang, Y., Shen, M.: Graph based Encrypted Malicious Traffic Detection with Hybrid Analysis of Multi-view Features. Information Sciences644, 119229 (2023)

work page 2023

[26] [26]

In: Proceedings of the 25th International Symposium on Research in Attacks, Intrusions and Defenses (RAID), pp

Fu, Z., Liu, M., Qin, Y., Zhang, J., Zou, Y., Yin, Q., Li, Q., Duan, H.: Encrypted Malware Traffic Detection via Graph-based Network Analysis. In: Proceedings of the 25th International Symposium on Research in Attacks, Intrusions and Defenses (RAID), pp. 495–509. ACM (2022)

work page 2022

[27] [27]

In: Proceedings of the 26th International Symposium on Research in Attacks, Intrusions and Defenses (RAID), pp

King, I.J., Shu, X., Jang, J., Eykholt, K., Lee, T., Huang, H.H.: EdgeTorrent: Real- time Temporal Graph Representations for Intrusion Detection. In: Proceedings of the 26th International Symposium on Research in Attacks, Intrusions and Defenses (RAID), pp. 77–91. ACM (2023)

work page 2023