acceptodds
Under review as a conference paper at ICLR 2027

AsymGAD: Label-Free Lateral-Movement Pivot Ranking via Asymmetric Graph Anomaly Detection

Abstract

Enterprise authentication activities naturally form graph-structured data that capture interactions among machines, users, and administrative services. Correspondingly, graph anomaly detection (GAD) provides a promising paradigm for identifying compromised machines involved in lateral movement, where attackers expand access by reusing legitimate credentials and services. However, detecting the intermediate stepping stones, or pivots, remains challenging because malicious authentication behavior often resembles routine activity, while the underlying interaction patterns continuously evolve over time. To address these challenges, we propose AsymGAD, a label-free GAD framework for ranking suspicious pivot machines in continuous authentication streams. AsymGAD first constructs evidence-sufficient observation graphs directly from the event stream, allowing graph boundaries to adapt to changing activity rather than relying on externally specified snapshots. It then introduces directional expansion as an explicit structural prior to capture machines whose outgoing behavior is disproportionately strong relative to their incoming activity, providing an interpretable signal of pivot behavior. Furthermore, AsymGAD incorporates causal information across consecutive observation graphs to distinguish persistent benign infrastructure from transient suspicious behavior, while a lightweight directed graph neural network captures complementary relational patterns beyond node-level statistics. By integrating adaptive graph construction, asymmetric structural cues, temporal context, and learned graph representations, AsymGAD enables label-free and interpretable pivot prioritization in large-scale authentication networks. Extensive experiments on a billion-event enterprise authentication stream demonstrate the effectiveness and scalability of AsymGAD for identifying compromised pivot machines under realistic large-scale conditions.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Networks, Anomaly Detection, Unsupervised Learning, Machine Learning.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_W3L9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ASYMGAD ranks potential lateral-movement pivot machines from enterprise authentication streams without attack labels. It combines adaptive window construction, a directional expansion score, historical persistence filtering, and a self-supervised directed GNN that predicts persistence percentiles. Evaluation on the LANL dataset reports improvements over seven baselines, with ablations identifying persistence filtering as the main contributor and message passing providing additional gains.

Critical strengths
  1. Practical and interpretable formulation. Authentication activity naturally forms a directed graph. The framework explicitly addresses observation-window construction and uses historical expansion behavior to suppress recurring high-fanout hosts. Its main scoring and filtering decisions are understandable and operate without incident labels.

  2. Useful experimental diagnostics. The count-matched random-removal control supports the conclusion that persistence filtering benefits from selecting particular hosts rather than merely reducing the candidate list. Component ablations and results across three training seeds also make the sources and variability of the improvements easier to assess.

Critical weaknesses
  1. Important comparisons are missing. Persistence filtering alone achieves 0.2402 mean AP, compared with 0.2616 for the complete method averaged over three seeds. Since the GNN predicts historical persistence, a direct persistence-based scoring baseline is needed to establish the benefit of learning this quantity from graph structure. Furthermore, the CybGFM comparison omits pretraining, limiting conclusions about superiority over the complete published method.

  2. Evidence of broad attack coverage remains limited. The evaluation contains only four pivot identities, with one appearing in 44 of 54 positive windows. The low-volume pivot C19932 has a best rank of 263, indicating limited recovery at practical analyst budgets. Table 9 reports window-level AP conditioned on a pivot’s presence, so improvements can arise from ranking another pivot in the same window. The hard persistence filter also suppresses sufficiently persistent attackers, restricting applicability to sustained compromises.

Questions for authors
  1. How does the GNN compare with directly incorporating historical persistence into the ranking under identical filtering and fusion conditions?
  2. Does the comparative advantage persist when CybGFM’s pretraining stage is reproduced?
  3. Can you report identity-specific Hit@K and results excluding the dominant pivot, separating low-volume detection from improvements attributable to co-occurring attacks?
3 comments

Response to Reviewer ---- Part 1/2

We thank the reviewer for the comments. We respond point by point with the additional experiments.

Comment 1.

"Persistence filtering alone achieves 0.2402 mean AP, compared with 0.2616 for the complete method averaged over three seeds. Since the GNN predicts historical persistence, a direct persistence-based scoring baseline is needed to establish the benefit of learning this quantity from graph structure."

"How does the GNN compare with directly incorporating historical persistence into the ranking under identical filtering and fusion conditions?"

Appendix D.11 fixes the windows, the filter and the fusion weight, and replaces the learned branch by the historical persistence ranking (Table 1). The 0.2402 value is the directional asymmetry score with the filter; the persistence ranking alone reaches 0.0252, the same ranking with the filter 0.1238, and the fused persistence order 0.2510. The complete method reaches 0.2616 as the three-seed mean and 0.2724 on the reported run, higher than the fused persistence order by 0.0106 and 0.0214, respectively.

We further scan K=10,20,30,50,100 and smoothed persistence. The best training-free configuration on LANL is P(K=20,N≥3) with a fused score of 0.2510; the default K=50 gives 0.2215, below the filtered asymmetry score of 0.2402. The learned branch is 0.0214 above the best training-free configuration on LANL. On OpTC the fused persistence order reaches 0.4647 while the learned branch reaches 0.7194, the mean of three runs. The branch prediction correlates 0.46 with the current-window asymmetry order and 0.20 with the persistence statistic: the learned output differs from the direct persistence order and adds a larger fused gain on OpTC. The K-scan is reported in Appendix D.11.

Table 1: Direct persistence ranking against the learned branch, mean AP.

ScoreLANLOpTC
Persistence alone0.02520.0299
Persistence with filter0.12380.0624
Persistence ranking, fused0.25100.4647
Learned branch, fused0.27240.7194
Random ranking0.000900.0308

The LANL learned-branch entry in Table 1 is the reported seed 0; its three-seed mean is 0.2616. The OpTC learned-branch entry is the mean of three runs.

Response to Reviewer ---- Part 2/2

Comment 3.

"Evidence of broad attack coverage remains limited. The evaluation contains only four pivot identities, with one appearing in 44 of 54 positive windows. The low-volume pivot C19932 has a best rank of 263, indicating limited recovery at practical analyst budgets. Table 9 reports window-level AP conditioned on a pivot's presence, so improvements can arise from ranking another pivot in the same window."

"Can you report identity-specific Hit@K and results excluding the dominant pivot, separating low-volume detection from improvements attributable to co-occurring attacks?"

We add identity-specific Hit@K, the evaluation without C17693, and the per-pivot decomposition of multi-pivot windows. In Appendix D.3 the mean AP is computed over the windows in which each pivot is active, while the rank and Hit@K refer to the host itself.

On the reported run, C17693 reaches 0.3336 mean AP with median rank 9 and best rank 1, and C22409 reaches 0.1597 with median rank 34. The identity-specific Hit@K is: C17693 hit@1 0.25, hit@10 0.52, hit@50 0.55, hit@100 0.59; C22409 hit@50 0.67, hit@100 0.67. The two low-volume pivots, C19932 and C18025, remain below practical analyst budgets up to K=100; their per-pivot statistics are reported in the same appendix.

Appendix D.13 evaluates the ten positive windows without C17693: our method reaches 0.00354, against 0.00091 for a random ranking and 0.00115 for the best baseline, CyberGFM; the three runs give 0.00354, 0.00365 and 0.00356. The relative advantage remains after removing the dominant pivot. In the sole-window analysis of the same appendix, C22409 reaches 0.0294 mean AP; the low-volume pivots remain limited.

Appendix D.14 decomposes the AP of each multi-pivot window into per-pivot contributions. In window 105 the AP is 0.1943, with C17693 contributing 0.1667 at rank 2; in window 113 the AP is 0.2502, with C17693 contributing 0.2500. The decomposition separates the contribution of a low-observation pivot from the gain of C17693 in the same window, so the evaluation of a single host is not replaced by the window-level AP.

Comment 4.

"The hard persistence filter also suppresses sufficiently persistent attackers, restricting applicability to sustained compromises."

Appendix D.7 quantifies how the persistence threshold affects pivot retention. The filter removes hosts whose observed persistence P exceeds Θ=2/3; the highest observed persistence among labeled pivots is 0.56, so the reported threshold retains all labeled pivots. Lowering Θ to 0.55 and 0.50 would remove 2 and 19 pivot records, respectively.

The method targets transient lateral movement. An attacker whose observed persistence exceeds the threshold is also removed. The Limitations appendix states the corresponding boundary: an adversary that holds a machine over many windows, or whose fanout is no higher than that of comparable hosts, may not be separated by either signal.

We will reflect these results and the scope statement in the revision, and thank the reviewer again.

Dear authors,

Thank you for the additional results. The comparison with pretrained CybGFM is still missing, and detection of low-volume pivots remains limited. I will keep my original rating of Weak Rejection.

Best regards, Reviewer W3L9

Reviewer_1kfVOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper discusses issues in graph anomaly detection. The main contribution of the paper is section 5 where it discusses LANL datasets and experiments.

It is difficult to ascertain the exact technical contribution from the allotted pages of the submission. The reviewer did notice a lot of added material in the appendices, including discussion or related work, tables of result, algorithm description. But the page allotment serves as a uniform baseline and it must be observed that the authors do not demonstrate comprehension of what constitutes a technical conference paper. Case in point, the figma/equivalent figure 1 is information free. It is a fantastic picture for a business or trade practice journal. Because of discussion like this, and the low information content of the equations of windowing etc., this document demonstrates that it does not have connection to technical audiences.

To ensure that a mistake is not being made, the algorithm 2 in appendix was read and it was not found to have novel content either. There are a lot of ablation experiments -- but ablation about what?

Critical strengths

The paper does state it clearly that selection of units and selection of boundaries is the critical part of anomaly detection. This is a strength, but that critique about units and boundaries has been well known for decades.

The paper does look into a number of recently published methods and experiments.

Critical weaknesses

The paper is incomprehensible in the pages it has put forward, and even the spot checked appendix does not answer the question about the novelty and ideas. Specific example: the phrase "enough events and enough distinct sources" was used multiple times in this paper; lines 73 (in promoting this document), lines 135 (in demoting prior work) and line 408 describing itself. But the description is vacuous --- "enough" is not a self-descriptive technical term.

The paper is not great in its own measurements; Table 1 shows a median average precision is 0.0082. And the range is 1.0 to 0, so there are no obvious scaling issues.

It is possible, and likely that the papers cited by this paper had poor experimental choices -- but pointing out mistakes (in not critical or well known papers) does not merit a publication in a technical venue. The authors always remain free to explore alternate mediums.

Questions for authors

None. This presentation is unsuitable for technical venues.

3 comments

Response to Reviewer ---- Part 1/2

We thank the reviewer for the detailed reading of the main text and the appendix. We respond to each point below.

Comment 1.

"The main contribution of the paper is section 5 where it discusses LANL datasets and experiments."

"It is difficult to ascertain the exact technical contribution from the allotted pages of the submission."

The technical contribution is developed in Section 4 and evaluated in Section 5. Section 4 defines the adaptive window rule (Eq. 1), the asymmetry score, the self-supervised structural branch trained on a label-free target, and the cross-window persistence memory with the fused score (Eqs. 2–3). Section 5 measures these components with controls: fixed grids for the window rule (Appendix D.1), a graph-free arm for the branch (Appendix D.5), a count-matched deletion for the filter (Appendix D.7), and a hand-set sweep for the fusion weight (Appendix D.6). The novelty is the combination of these four components under one label-free protocol: boundaries are chosen online from the stream, the branch is trained on a label-free target, and the filter and the fusion weight are computed from window statistics (Sections 4.4–4.6).

Comment 2.

"Because of discussion like this, and the low information content of the equations of windowing etc., this document demonstrates that it does not have connection to technical audiences."

"Specific example: the phrase "enough events and enough distinct sources" was used multiple times in this paper; lines 73 (in promoting this document), lines 135 (in demoting prior work) and line 408 describing itself. But the description is vacuous --- "enough" is not a self-descriptive technical term."

As the reviewer notes, "enough" is not a self-descriptive technical term, and the revision does not retain the phrase. Eq. (1) states the rule with four quantities: the numbers of raw events, directed pairs, distinct sources and distinct destinations; a window closes at the first event at which all four hold, and the values are fixed before the stream is read and listed in Appendix C.2. AWC serves two purposes: reducing the dilution of the attack signal by normal activity, and providing a stable supply of information to the learned component. On LANL it maintains about 2×10^6 events per window, with a standard deviation of 2.2% of the mean. At matched volume the filtered score reaches 0.2402 with AWC, against 0.2130 for fixed event-count windows (12.8%) and 0.2096 for wall-clock windows (14.6%) (Appendix D.1).

Comment 3.

"To ensure that a mistake is not being made, the algorithm 2 in appendix was read and it was not found to have novel content either."

Algorithm 2 ranks the candidates of one window and updates the persistence state. Three of its design choices are specific to this setting. The refit counter and the training buffer are positions in the stream: the branch is refitted every eight observation windows on the sixty most recent windows, and a fit draws on windows up to the one being scored (Section 4.4). The persistence memory is read at the start of a ranking and updated at its end; it holds the fraction of earlier windows in which a host entered the extreme-asymmetry set (Section 4.5). The fusion weight is computed from the spread statistic of the current window (Section 4.6). Appendices D.4–D.7 quantify the components: the filter raises mean AP from 0.0858 to 0.2402; a count-matched random deletion leaves 0.0862; the branch adds 0.0322; a graph-free arm adds 0.0040; and the adaptive weight is within 0.001 of the best hand-set value.

Comment 4.

"There are a lot of ablation experiments -- but ablation about what?"

Each comparison in Section 5.5 and Appendix D isolates one factor of the final score while the rest of the pipeline is fixed. The component ablation separates the unlearned score, the branch and the filter (0.0858, 0.1294, 0.2402, 0.2724; Appendix D.4). A count-matched random deletion leaves 0.0862, which attributes the filter's gain to the hosts it removes (Appendix D.7). A graph-free arm adds 0.0040, which attributes the branch's 0.0322 increment to message passing (Appendix D.5). The hand-set sweep tests the fusion weight (Appendix D.6), the fixed grids test the window rule (Appendix D.1), and three runs test run-to-run variation (0.2616 ± 0.0111; Appendix D.9). The revision states this mapping at the start of Section 5.5.

Response to Reviewer ---- Part 2/2

Comment 5.

"The paper is not great in its own measurements; Table 1 shows a median average precision is 0.0082. And the range is 1.0 to 0, so there are no obvious scaling issues."

Table 1 reports both the mean and the median over the 54 positive windows. A known pivot ranked at the head of a window contributes an AP close to one, and the value decreases with the rank of the pivot; the mean reflects the recovered windows and the median the typical window. The two statistics agree in their ordering: the median of AsymGAD is 0.0082, the highest among the methods in Table 1 (the best baseline gives 0.0072), and the mean is 0.2724 against 0.0299. Of the 54 windows, 27 have AP at least 0.01 (at most 17 for any other method), 28 give AsymGAD a unique win (at most 13 for any other), and it places a known pivot in the top 10, 50 and 100 in 23, 25 and 27 windows (18, 24 and 25 for the unlearned score); every paired bootstrap interval excludes zero in all three runs (Appendix D.2; Appendix D.3).

Comment 6.

"It is possible, and likely that the papers cited by this paper had poor experimental choices -- but pointing out mistakes (in not critical or well known papers) does not merit a publication in a technical venue. The authors always remain free to explore alternate mediums."

The contribution of the paper is the detector of Section 4 and its evaluation in Section 5; the comparison with prior work motivates the evaluation protocol. Published detectors are typically evaluated on graphs whose boundaries are given in advance, and our setting constructs the boundaries online and without labels (Section 2). Each method in the benchmark is evaluated on the same graphs, candidate sets and ranking code, with the anomaly direction and normalization of its original paper (Section 5.1). The framework is also constructive: Appendix D.10 applies the asymmetry score and the persistence filter to seven public baselines, and each baseline improves over its own re-run value (AnomalyGFM 0.0101 to 0.3079; CyberGFM 0.0071 to 0.2703).

Comment 7.

*"The paper does state it clearly that selection of units and selection of boundaries is the critical part of anomaly detection.

This is a strength, but that critique about units and boundaries has been well known for decades."* The observation that the unit of observation affects detection is long established and serves as the motivation for the protocol. It appears in the statistical intrusion-detection model [1], in network monitoring where the window length and failure threshold are described as "probably the most difficult to set a priori" [2], in multi-timescale traffic analysis where filters tuned to specific aggregation levels expose different anomaly characteristics [3], and it remains an active empirical question [4]. The paper's contribution is the use of the observation unit as part of the detector: the boundary rule closes a window on four label-free evidence thresholds (Section 4.2), and the effect of the choice is measured with the scoring pipeline held fixed (Appendix D.1). Across six segmentations the filtered score ranges from 0.1925 to 0.2402, and at matched volume the adaptive rule reaches 0.2402 against 0.2130 for fixed event-count windows and 0.2096 for wall-clock windows (Appendix D.1).

Comment 8.

"The paper is incomprehensible in the pages it has put forward, and even the spot checked appendix does not answer the question about the novelty and ideas."

"Presentation: 1; Informativeness: 1; Soundness: 1. This presentation is unsuitable for technical venues."

The revision makes the technical content easier to locate: Section 4 states the method and its equations, Section 5 measures each component with a control, Section 4.2 defines the boundary rule with four evidence thresholds, Appendix C.1 states the three label-free constraints of Algorithm 2, Section 5.5 maps the ablation comparisons, and Section 5.2 reads Table 1 by mean and median. The overview figure is a schematic of the pipeline; the technical content is in Equations (1)–(5), Algorithms 1–2 and the appendix tables. The evaluation is bounded by the public data: four pivot identities on one enterprise authentication stream, with the limits stated in Limitations.

References

[1] D. E. Denning. An Intrusion-Detection Model. IEEE Transactions on Software Engineering, 1987.

[2] J. D. Brutlag. Aberrant Behavior Detection in Time Series for Network Monitoring. LISA, 2000.

[3] P. Barford et al. A Signal Analysis of Network Traffic Anomalies. ACM SIGCOMM IMW, 2002.

[4] A. Y. J. Yang et al. Graph-Based Anomaly Detection in LTE Networks: Effects of Temporal Aggregation and Window Alignment. IEEE ISNCC, 2026.

We thank the authors for commenting on the comments.

The "main" contribution is indeed section 5. Section 4 has design rationale (4.1), an adaptive window construction(with acronym AWC, but all it is doing is making sure that the number of events is M_{min} and then similar lower bounds on source destination), 4.3 which uses a lot of words to say that a graph will be constructed and the "features" are the in and out degree, 4.4 layer definition (this section can have some content), 4.5 (percentile rank of an edge in its two endpoints), 4.6 something called Adaptive fusion; the words adaptive fusion has been used twice in this manuscript, it is used in Figure 1, literal white box, and never defined. The formulas in the rest of the section are ratios of mean/median without any discussion of why anyone should care. There is a lot of text in section 4; it is possible that authors feel that those are contribution. This review does not view these as contribution. The entire section 4 can be compressed in a few lines instead of a framework embellishment.

The fact that anyone should care is in section 5, assuming the results are stellar. They did not seem to be.

All of the equations and pointers can then be added easily to "In summary, this work makes four contributions...."

This is not about "and the revision does not retain the phrase"> The paper needs to be written from scratch with a purpose about what it want the reviewer to know -- in the current presentation that purpose is carried by Figure 1, a system diagram which is not a scientific contribution that this review is set up to assess. We are however happy that the authors realize that the phrasing of this text needs a revision.

We also thank the authors for accepting the review "The observation that the unit of observation affects detection is long established and serves as the motivation for the protocol" in their next message. That removes any claim that the problem formulation is new. But protocol? The stated goal of this paper as written is to detect plausibility of lateral movement; from a technical standpoint the paper will be evaluated on that task alone. The reviewer recommends reducing the use of protocol/framework similar, terms and provide technical specifics about why various choices matter.

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

The authors propose a method for identifying potentially compromised computers within corporate networks by analyzing authentication patterns between machines. It addresses the unsupervised identification and ranking of hosts involved in lateral movement, which is a big problem in this field of research.

The main idea is relatively simple and intuitive. A computer used by an attacker to access other machines may exhibit a connection pattern that differs from normal network behavior and the proposed method combines a heuristic based on this behavior, a temporal memory mechanism, and a GNN to generate a ranking of the most suspicious computers.

The authors experimented with the LANL dataset and presented results with higher average precision than the evaluated baselines, which is very promising, but the gains are concentrated in the persistence filter and a small subset of attack windows, which suggests new investigation is needed.

Critical strengths

This is one of the more relevant problems in practical cybersecurity operations and being able to rank suspicious hosts is very useful for prioritizing investigations under limited analyst capacity or a high volume of data.

I found it very interesting that the formulation considers the temporal nature of authentication events and explicitly avoids using attack labels during model inference because this is one of the big problems with cybersecurity models approach, the need for labeled data in an environment where zero-day discovery is the target.

The paper is clear and the experimental reporting is easy to understand as the authors used per-window experiments, parameter sensitivity, random seeds, and ablation analysis. It allows the reader to have an independent verification of the experimental setup and results.

I liked that the authors included useful controls to separate the effects of persistence filtering, graph message passing, and window construction. This is an important step that is normally neglected in papers in this field. The use of ranking metrics and analyst-oriented measures is appropriate for the context of this application.

Critical weaknesses

Besides the interesting approach, I think the work does not yet demonstrate that the gains stem from the proposed learning architecture, or even that they generalize to different attacks and networks. If I understand correctly, the results appear to be explained primarily by a temporal filtering heuristic and by that I don't mean the heuristic is useless. On the contrary, it could be a simple and efficient tool. But I think the evidence supports a more modest contribution than the one presented by the authors.

For instance, the experimental evidence is dominated by a single attacking host. C17693 appears in 44 of the 54 positive windows. And 11 windows account for approximately 75% of the reported mAP. I think this is limiting the evidence of generalization across different attack patterns and the authors should comment more on that.

I found that the principal improvement comes from a simple persistence filter rather than the learned component. Please correct me if I am wrong, but it seems filtering alone increases mAP from approximately 0.086 to 0.240, while the graph-based component produces comparatively small and seed-dependent improvements.

I think some baselines perform below the expected random ranking level, consider the case of the CybGFM method, it is evaluated without its pretraining stage, and the proposed persistence filter is not applied to competing methods.

This point needs to be better explained, I think several design decisions are supported by statistics from the labeled evaluation windows. In this case, without a separate validation protocol, it is difficult to establish that the method and its thresholds generalize beyond the evaluated attack.

Please comment on the fact that the adaptive window mechanism seems not to be distinguished from fixed event-count windows, and the adaptive fusion weight provides little measurable improvement over a fixed value.

The evaluation protocol needs to be expanded, as it is restricted to LANL and without comparison against relevant host-oriented lateral movement detection methods or another independent dataset, it is not easy to evaluate the models' effectiveness.

Questions for authors

I think the reader will benefit from some guidance from the authors in relation to those points:

  • What is the performance of each baseline when the same persistence filter is applied?
  • It would be very interesting to include simple baselines based on out-degree, directional asymmetry, and historical frequency in the experiments.
  • How do the authors explain the below-random performance observed for some baselines?
  • Were the ranking direction and score normalization verified?
  • What are the results when C17693 is excluded from the evaluation? This is a very interesting scenario to evaluate.
  • Does the method retain meaningful detection performance on the remaining pivots?
  • How were the persistence threshold, window parameters, and other constants selected?
  • Were labeled evaluation windows used during method development?
  • Why were additional datasets and closely related approaches, such as Argus, Hopper, and LARES, not included in the experimental setup?
  • Please comment on how the window-level average precision can reflect the detection of another pivot present in the same window.
4 comments

Response to Reviewer ---- Part 1/4

We thank the reviewer. Our work is framed by two constraints: the application is lateral movement inside an enterprise network, and the method must operate without attack labels. These constraints drive the dataset choice, the experimental protocol and the design; we explain each point in turn.

Comment 1.

"The evaluation protocol needs to be expanded, as it is restricted to LANL and without comparison against relevant host-oriented lateral movement detection methods or another independent dataset, it is not easy to evaluate the models' effectiveness."

"Why were additional datasets and closely related approaches, such as Argus, Hopper, and LARES, not included in the experimental setup?"

We surveyed LANL, OpTC, UNSW-NB15, DAPT and IoT-23. UNSW-NB15 targets network-intrusion traffic [1]; DAPT2020 provides an APT chain with a lateral-movement stage but labels each flow on a six-host internal network with a single positive host [2]; IoT-23 targets IoT malware traffic. UNSW-NB15 and IoT-23 lack internal host relations, and DAPT2020's host population is too small to serve as the main evaluation corpus for our large-scale host-ranking task; Hopper's enterprise login data are not public [3]. In our survey, LANL [4] is the only public dataset that combines a large-scale enterprise authentication graph, verifiable lateral-movement labels and temporal separation [5]. OpTC can be partially adapted: on its host-to-host graph the full method reaches 0.7194 mean AP and places a known pivot in the top ten in all ten windows, tied with GAD-NR and ahead of the other six baselines (Appendix D.15). A host-level port of Argus on LANL gives our method 0.2724 mean AP against 0.1549 over the 54 windows, and 0.3645 against 0.1782 over the 33 windows in the first 14 days (Appendix D.16). LARES's released protocol differs; the protocol comparison is in Comment 3, and Hopper releases no host-level detector [3].

Comment 2.

"For instance, the experimental evidence is dominated by a single attacking host. C17693 appears in 44 of the 54 positive windows. And 11 windows account for approximately 75% of the reported mAP. I think this is limiting the evidence of generalization across different attack patterns and the authors should comment more on that."

"What are the results when C17693 is excluded from the evaluation? This is a very interesting scenario to evaluate."

"Does the method retain meaningful detection performance on the remaining pivots?"

The main difficulty on LANL is that malicious hosts are extremely few and highly imbalanced: four labeled pivots, while a window has a median of 9,780 active hosts. C17693 is active in 44 of the 54 positive windows, and, as you note, about 75% of the reported mAP comes from 11 windows; a small number of hosts and windows therefore dominates the aggregate. We report results per pivot (Appendix D.3) and add an evaluation without C17693 (Appendix D.13) to make clear what the aggregate reflects and where its limits are. The main case host 9660 of LARES is the same host as our C17693 [6], which reflects how few lateral-movement samples the public data offer. On the ten positive windows without C17693, AsymGAD reaches 0.00354 against 0.00091 for a random ranking and 0.00115 for the best baseline, CyberGFM (Appendix D.13); the relative advantage remains, and the absolute level is bounded by the few positives available on the remaining pivots.

Response to Reviewer ---- Part 2/4

Comment 3.

"This point needs to be better explained, I think several design decisions are supported by statistics from the labeled evaluation windows. In this case, without a separate validation protocol, it is difficult to establish that the method and its thresholds generalize beyond the evaluated attack."

"How were the persistence threshold, window parameters, and other constants selected?"

"Were labeled evaluation windows used during method development?"

The main difficulty of this scenario is that malicious hosts are extremely few relative to the candidate set and that normal activity forms a background for the attack signal. Data scope, event filtering, window construction and aggregation granularity directly change the candidates that enter the evaluation and the behavioral evidence available; we therefore treat them as part of the protocol and report them with the performance results. Studies of experimental bias and data leakage show that dataset construction and temporal splitting can affect reported machine-learning performance [7, 8, 9].

Our protocol constraints are concrete. Window boundaries are closed online on four label-free evidence thresholds; the training buffer and the refit cadence are positions in the stream, fixed before the stream is read; the constants are set from label-free quantities (Appendix C.2 and E); the 54 labeled windows serve only the final metrics and take no part in any design choice; and the sensitivity of the thresholds and the fusion weight is reported in Appendix D.6 and D.7. The design decisions are therefore not supported by statistics from the labeled evaluation windows.

Existing work adopts different operating assumptions. Argus uses the first attack time as its training cut-off, and the LANL train, validation and test boundaries of LARES are placed relative to the first attack time [6, 10]; these settings are compatible with their unsupervised losses, but their LANL experimental splits use the first-attack time information. Our setting corresponds to a deployment in which the attack start is unknown and labels cannot be revisited, so we require training intervals, online window boundaries and parameters to use no attack labels.

We therefore emphasize the task conditions behind different protocols and their comparability. The changes in baseline levels in the supplementary experiments come mainly from corpus scale and label density; the comparisons in the paper are made within each protocol, and OpTC is reported as a restricted-protocol partial-mechanism check.

Comment 4.

"Please comment on the fact that the adaptive window mechanism seems not to be distinguished from fixed event-count windows, and the adaptive fusion weight provides little measurable improvement over a fixed value."

AWC serves two goals: reduce the dilution of the attack signal by normal activity, and give AsymGAD a stable supply of information. AWC closes windows online on evidence thresholds; on LANL it holds about 2M events per window with a standard deviation of 2.2% of the mean. At matched volume, A+P reaches 0.2402 with AWC against 0.2130 for fixed event-count windows, a 12.8% improvement; Table 1 gives the full comparison, including fixed wall-clock windows. The gain has a clear boundary: in 523 of the 525 windows both schemes close at the same amount of evidence, and the difference concentrates in the few windows with abrupt traffic changes, so we read the adaptive rule as a supplement in those windows. Parameters depend on data scale and activity patterns; our setting is a reference configuration, chosen from data-scale priors without label tuning, and sensitivity to multiple parameter settings is in Appendix D.6 and D.7. The fusion weight is fixed from priors rather than adapted per window; the sensitivity curve is flat, and the best hand-set value differs by less than 0.001 mean AP (Appendix D.6), so adaptive weighting offers little measurable improvement and the reported results use the fixed value.

Table 1: A+P under different observation-window schemes on LANL.

SegmentationA aloneA+P
AWC0.08580.2402
Wall clock, 1 h0.07390.1925
Wall clock, 2 h0.08850.2331
Wall clock, 3 h0.07520.2096
Event count, 2M0.07810.2130
Event count, 3M0.07460.2338

Response to Reviewer ---- Part 3/4

Comment 5.

"Besides the interesting approach, I think the work does not yet demonstrate that the gains stem from the proposed learning architecture, or even that they generalize to different attacks and networks. ... But I think the evidence supports a more modest contribution than the one presented by the authors."

"I found that the principal improvement comes from a simple persistence filter rather than the learned component. Please correct me if I am wrong, but it seems filtering alone increases mAP from approximately 0.086 to 0.240, while the graph-based component produces comparatively small and seed-dependent improvements."

The filter is strong because node roles differ sharply in enterprise networks. Our LANL statistics show the top of the raw ranking is dominated by infrastructure such as scanners, backup and patch servers: the filtered hosts have a mean in-degree of 0.8 and initiate 99.9% of their contacts. These hosts are highly active but are not lateral movement. Under the reported configuration the filter removes about 19–23 hosts per window and no labeled pivot (Appendix D.7). This supports the design choice: node-local directional evidence and scenario priors narrow the candidate range first, and graph correlation supplements it when node evidence is insufficient. The filter handles clear infrastructure behavior directly and therefore yields a large gain; the learned component adds mainly on low-observation pivots where node evidence is ambiguous, with gains of 0.0322, 0.0219 and 0.0100 over three runs (Appendix D.3, D.9).

To test whether the branch merely re-learns the statistic it is trained on, we replace it by that statistic (Table 2): the replacement reaches 0.2510 on LANL, but on OpTC it drops the fused score from 0.6194 to 0.4647 while the branch raises it to 0.7194 (Appendix D.11). The branch therefore carries structure information that the statistic does not, and the pattern is consistent with its gains concentrating where node evidence is thin.

Table 2: Mean AP decomposition on LANL.

ConfigurationMean AP
A alone0.0858
A+P0.2402
Full method0.2724
Persistence ranking instead of branch0.2510

Comment 6.

"I think some baselines perform below the expected random ranking level, consider the case of the CybGFM method, it is evaluated without its pretraining stage, and the proposed persistence filter is not applied to competing methods."

"What is the performance of each baseline when the same persistence filter is applied?"

"How do the authors explain the below-random performance observed for some baselines?"

"Were the ranking direction and score normalization verified?"

We compare the seven baselines alone, with A, and with A+P (Appendix D.10). With A alone every baseline improves by 3.3× to 57× relative to its re-run value; with the filter added, all seven exceed the unprocessed re-run value, and AnomalyGFM and CyberGFM reach 0.3079 and 0.2703 (Table 3). These results show that the framework can be combined with other anomaly detectors. Methods with graph pretraining or neighborhood propagation keep their structural representation while receiving A's ranking signal and the persistence constraint, which is why they complement the framework most strongly. On this protocol, several baselines fall below the random reference of 0.00090 (Table 4); the values are computed under the unified candidate sets. All baselines are processed with the anomaly direction and normalization of their original papers, and AP is computed on the final ranking. The CybGFM reproduction is discussed in Comment 8.

Table 3: The seven public baselines with A and A+P injected, mean AP.

Methodalone+A+A+P
GAD-NR0.00030.01710.0246
GEL0.00130.03160.0778
TAM0.00030.00180.0018
AnomalyGFM0.01010.09110.3079
EULER0.00460.05580.0622
CybGFM0.00710.11620.2703
CoLA0.02050.06810.1731

Response to Reviewer ---- Part 4/4

Comment 7.

"It would be very interesting to include simple baselines based on out-degree, directional asymmetry, and historical frequency in the experiments."

"Please comment on how the window-level average precision can reflect the detection of another pivot present in the same window."

Appendix D.12 reports the simple host-level scores (Table 4). These results show that simple host-level statistics are strong in their own right. Window-level AP reflects every pivot in the window. Appendix D.14 gives the decomposition for all seven multi-pivot windows.

Table 4: Simple host-level scores, mean AP.

ScoreUnfilteredWith filter
Out-degree0.04240.1069
Directional asymmetry0.08580.2402
Historical frequency0.00280.0028
Random0.000900.00090

Comment 8.

"consider the case of the CybGFM method, it is evaluated without its pretraining stage."

We also state the limits of reproduction. The public CybGFM code provides no pretrained weights; with the public implementation we cannot guarantee the best performance of the original method. Except for the omitted pretraining, we keep its original settings and add no extra tuning, and mark it as a pretraining-free port. These results represent this pretraining-free port; the effect of full pretrained weights needs further verification.

We also compare the method with a host-oriented detector. Table 5 reports a host-level port of Argus on LANL (Appendix D.16): our mean AP is 0.2724 against 0.1549 over the 54 positive windows, and 0.3645 against 0.1782 over the 33 windows in the first 14 days.

Table 5: Host-level comparison with Argus on LANL.

MetricOursArgus
Mean AP, 54 windows0.27240.1549
Mean AP, first 14 days0.36450.1782

We will reflect these additions in the revised version, and thank the reviewer again.

References

[1] N. Moustafa and J. Slay. UNSW-NB15: a comprehensive data set for network intrusion detection systems (UNSW-NB15 network data set). In 2015 Military Communications and Information Systems Conference (MilCIS), pp. 1-6, 2015.

[2] S. Myneni, A. Chowdhary, A. Sabur, S. Sengupta, G. Agrawal, D. Huang, and M. Kang. DAPT 2020 - Constructing a Benchmark Dataset for Advanced Persistent Threats. In Deployable Machine Learning for Security Defense (Communications in Computer and Information Science), pp. 138-163. Springer, 2020.

[3] G. Ho, M. Dhiman, D. Akhawe, V. Paxson, S. Savage, G. M. Voelker, and D. Wagner. Hopper: Modeling and Detecting Lateral Movement. In 30th USENIX Security Symposium (USENIX Security 21), pp. 3093-3110, 2021.

[4] A. D. Kent. Cyber Security Data Sources for Dynamic Network Research. In Dynamic Networks and Cyber-Security, chapter 2, pp. 37-65. World Scientific, 2015.

[5] C. Smiliotopoulos, G. Kambourakis, and C. Kolias. Detecting Lateral Movement: A Systematic Survey. Heliyon, 10(4):e26317, 2024.

[6] T. Bilot, N. El Madhoun, A. Zouaoui, K. Al Agha, and T. Pasquier. LARES: Host-Centered Lateral Movement Detection via Inductive Graph Reasoning. In IEEE Annual Computer Security Applications Conference (ACSAC 2026), 2026. To appear.

[7] D. Arp, E. Quiring, F. Pendlebury, A. Warnecke, F. Pierazzi, C. Wressnegger, L. Cavallaro, and K. Rieck. Dos and Don'ts of Machine Learning in Computer Security. In 31st USENIX Security Symposium (USENIX Security 22), pp. 3971-3988, 2022.

[8] F. Pendlebury, F. Pierazzi, R. Jordaney, J. Kinder, and L. Cavallaro. TESSERACT: Eliminating Experimental Bias in Malware Classification across Space and Time. In 28th USENIX Security Symposium (USENIX Security 19), pp. 729-746, 2019.

[9] M. A. Bouke and A. Abdullah. An Empirical Study of Pattern Leakage Impact During Data Preprocessing on Machine Learning-Based Intrusion Detection Models Reliability. Expert Systems with Applications, 230:120715, 2023.

[10] J. Xu, X. Shu, and Z. Li. Understanding and Bridging the Gap Between Unsupervised Network Representation Learning and Security Analytics. In 2024 IEEE Symposium on Security and Privacy (SP), pp. 3590-3608, 2024.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.