On Fair and Realistic Performance Evaluations for Graph-Based Lateral Movement Detectors Corentin Larroche
arXiv:2607.29390v1 [cs.CR] 31 Jul 2026
French National Cybersecurity Agency (ANSSI), Paris, France [email protected]
Abstract. Research on lateral movement detection has made significant progress in recent years, spurred by the widespread availability of benchmark datasets that make evaluating detectors practical. However, the exact way in which these benchmark datasets are used varies across the literature: both the preprocessing applied before feeding the data to the detector and the labeling of lateral movement-related events change substantially from one paper to another. We survey preprocessing and labeling methodologies for two popular datasets and demonstrate their impact on the fairness and realism of downstream evaluations. We also propose well-grounded preprocessing and labeling policies for these datasets. Finally, we re-evaluate three widely cited lateral movement detection methods under these new policies; our results differ significantly from those reported in the original papers, further highlighting the critical importance of dataset preprocessing and labeling practices in evaluating lateral movement detectors. Keywords: Intrusion detection · Lateral movement · Evaluation.
1
Introduction
Benchmarking has long been a key accelerator of research in machine learning in general, and on applications of machine learning for intrusion detection in particular. The availability of widely shared evaluation datasets enables fair, systematic comparisons between competing models and algorithms, allowing researchers to identify the most promising approaches and easily assess their own ideas. This dynamic has played out over the last ten years in the subfield of lateral movement detection, where the goal is to identify connections within an enterprise network that are made by an attacker hopping from already compromised computers to new targets. Datasets such as LANL’s "Comprehensive, multi-source cyber-security events" [13,12] and DARPA’s "Operationally transparent cyber" [8] (OpTC) have been used by many researchers to develop new detectors and advance the state of the art. These two datasets consist of network and system logs from enterprise networks with traces of malicious activity performed by red teams, including lateral movements. This makes them good benchmarks for evaluating lateral movement detectors. However, actual detectors do not directly take raw events as input: some preprocessing is usually involved in turning these events into higher-level, typically
2
C. Larroche
graph-based representations. In addition, evaluating detectors requires precisely labeling events related to lateral movements. This is not straightforward, especially for the OpTC dataset, which has no event-level labels at all. Both preprocessing and labeling methodologies significantly impact downstream evaluations. More specifically, some choices lead to unrealistic evaluation settings: for instance, arbitrarily filtering out some events or labeling benign events as malicious can artificially inflate detection performance. Moreover, evaluating competing detectors under different preprocessing and labeling policies leads to unfair comparisons. Reliable evaluations thus require widely shared, well-grounded preprocessing and labeling methodologies. In practice, existing lateral movement detection research applies heterogeneous and sometimes questionable preprocessing and labeling to the LANL and OpTC datasets. We survey the evaluation settings found in the literature for these two datasets and highlight their diversity as well as their shortcomings with respect to realism. We also propose preprocessing and labeling policies justified by operational considerations and release code implementing them1 . Finally, we re-evaluate three previously published lateral movement detection methods under our proposed methodology and show that the obtained performance metrics differ significantly from those originally reported. The rest of this paper is structured as follows. We provide necessary background on lateral movement detection in Section 2. Existing preprocessing and labeling choices for the LANL and OpTC datasets are then reviewed in Section 3. Sections 4 and 5 focus more specifically on dataset preprocessing and labeling, respectively. Finally, our experimental results are presented in Section 6.
2
Background
We begin with an overview of important concepts and related work, starting with a definition of lateral movement detection in Section 2.1, then reviewing existing graph-based detection methods in Section 2.2, and finally focusing on how these methods are evaluated in Section 2.3. 2.1
Lateral Movement Detection
Lateral movement [1] is a popular offensive tactic as well as a critical step of advanced intrusions. It consists in leveraging an already compromised host to gain access to another host within the same local network. Typically, after obtaining an initial code execution on some host through social engineering, reuse of stolen credentials, or vulnerability exploitation, attackers move laterally through the local network to find more valuable hosts and data. In practice, many tools and techniques can be harnessed for lateral movement, and many of them rely on legitimate software (i.e., so-called livingoff-the-land binaries) and protocols. This includes software used by administrators for remote management, such as the Windows Management Instrumentation 1
https://github.com/cl-anssi/LMDEval/
time
Performance Evaluations for Graph-Based Lateral Movement Detectors A
B
A
C
B
C
A
C
B
E
A
D
B
E
A
C
B
E
Static
Continuous time
A
t2
E
C
Δ
3
B
C t4 t8
A t6
D
t1
t5 t3
B
E t7
t9
D
Discrete time (snapshot length Δ)
Δ A
t1
E
C
C B
A
B D
t2
Fig. 1. Three ways to turn a sequence of timestamped events into a graph. The width of each edge depends on the number of underlying events.
(WMI), and network protocols designed to manage sessions on remote computers (e.g., RDP or SSH). Both the diversity of methods and the opportunity for attackers to leverage legitimate, frequently used software and protocols make rule-based detection of lateral movement difficult: many rules must be created to cover various tools and techniques, and some of these rules might generate many false positives in real-world enterprise networks due to the presence of benign remote management activity. Anomaly detection thus appears as an appealing alternative to rule-based systems: instead of trying to precisely describe malicious lateral movement behavior, defenders can build a profile of legitimate usage of commonly weaponized tools and protocols within a given enterprise network. Such a profile is typically obtained by analyzing data collected during a training period. Malicious behaviors then stick out as significant deviations from this norm. In practice, though, anomaly-based intrusion detection tends to generate many false positives due to rare but legitimate behaviors, and lateral movement detection thus remains an active and challenging research topic. From a mathematical perspective, lateral movements can be characterized as anomalous edges in a directed graph whose nodes are the hosts of the local network, with each edge representing an information flow from one host to another. Such a graph is typically built using remote authentication logs or network flow records. Most research on lateral movement detection then focuses on applying graph-based anomaly detection techniques to this graph. 2.2
Existing Graph-Based Detection Methods
There are three main ways to build graphs from authentication logs or network flows, as illustrated in Figure 1: static graphs that aggregate all available data into a single instance, discrete-time graphs that divide the data into fixedlength windows and build one snapshot graph per window, and continuoustime graphs, where each edge is marked with a timestamp. Choosing one of these approaches implies a trade-off between the faithfulness of the representa-
4
C. Larroche
tion to the actual data and the complexity of the anomaly detection task. While excessively simple representations can destroy relevant information and thus miss some anomalies, an increase in complexity can lead to a higher computational cost and more false positives because of the added noise. Static graph-based methods usually represent the whole training period as a single graph, then use this graph to independently compute an anomaly score for each edge from the testing period. They rely on tools such as node embedding algorithms [28,6] (e.g., node2vec [9]), graph neural networks (GNNs [7]) [19,25], and community detection [26]. Discrete-time graph-based methods cut both the training and testing periods into fixed-size, disjoint snapshots. One graph is then built using the data from each snapshot. Finally, a model is trained on the sequence of training snapshots, then evaluated on the sequence of testing snapshots. Models used in this setting include dynamic latent space network models [18] and GNNs coupled with recurrent neural networks (RNNs) [16,17,30]. Finally, continuous-time graph-based methods preserve the exact timestamp of each event, representing the training and testing periods as sequences of timestamped edges. Various techniques can then detect anomalies within the testing sequence, from heuristics and simple statistics [20,11] to temporal graph embedding algorithms [31,23] and temporal graph networks (TGNs [24]) [14].
2.3
Evaluating Detectors: Datasets, Metrics, and Pitfalls
The two most popular datasets in lateral movement detection research are the "Comprehensive, multi-source cyber-security events" dataset [12,13] released in 2015 by the Los Alamos National Laboratory (referred to as the LANL dataset in the rest of this paper) and the "Operationally transparent cyber" dataset [8] released in 2020 by DARPA (here called the OpTC dataset). The LANL dataset contains multiple types of logs obtained from a real-world enterprise network with approximately 16,000 hosts over 58 days. The authentication logs and network flows from this dataset have been used to evaluate lateral movement detectors. Some labeled authentication events correspond to offensive actions performed by a red team, and these labels serve as a ground truth for lateral movements. As for the OpTC dataset, it contains system and network logs from a simulated enterprise network consisting of 1,000 workstations and a few servers. The network logs only record flows between internal and external hosts; they are thus not suited for lateral movement detection since lateral movements are, by definition, connections between internal hosts. Therefore, most research relies on the network-related events from the system logs instead. Note, however, that system logs are provided only for 500 workstations, preventing reconstruction of the whole flow history. In addition to synthetic background activity spanning 9 days, the OpTC dataset records three attack scenarios performed by a red team over the last three days. However, the events related to red team activity are not labeled; only a high-level report is available for each scenario. This raises some challenges when evaluating lateral movement detectors, which we further discuss in Section 5.3.
Performance Evaluations for Graph-Based Lateral Movement Detectors
5
Other datasets appear in the lateral movement detection literature, such as Pivoting [2] and PicoDomain [10]. Pivoting contains network flows from a real-world network but no actual malicious lateral movements; only legitimate pivoting activity is labeled. As for PicoDomain, it is a synthetic dataset based on a very small network (only six hosts) and thus cannot measure the reliability of a detector when deployed in a large real-world enterprise network. Therefore, we focus on the LANL and OpTC datasets in the rest of this paper. In addition to benchmark datasets, evaluating lateral movement detectors requires suitable performance metrics. Assuming that detectors return binary predictions (normal or anomalous event), the simplest of these metrics are the true positive rate (TPR, also named recall) and false positive rate (FPR). Since there are usually much fewer lateral movement-related events than benign ones, even a low FPR can lead to a large number of false positives drowning out the true positives (a problem sometimes referred to as the base rate fallacy [3]). Precision, defined as the ratio of true positives among all predicted anomalies, is thus more informative. The F1 score, defined as the harmonic mean of precision and recall, provides a global assessment of the ability of a detector to flag actual lateral movements without raising too many false alarms. In practice, however, most actual detectors return a continuous anomaly score rather than a binary prediction, and computing the aforementioned metrics thus requires setting a decision threshold. Alternative metrics avoid making this sensitive decision by computing averages over all possible thresholds: the area under the receiver operating characteristic (AUC) is defined as the average TPR over all possible values of the FPR, while the average precision (AP) is the mean of the precision over all possible values of the recall. We use the AUC and AP as performance metrics in the following sections. Given a dataset and a set of performance metrics, evaluating lateral movement detectors can appear as a straightforward task. In practice, existing replication studies [5,27] have uncovered several problems with the empirical evaluations performed in the lateral movement detection literature. These include insufficiently documented hyperparameter tuning, errors in the published implementations of the detectors, and poor generalization to new datasets. We address another issue that was left aside in these studies, namely dataset preprocessing and labeling. The core of this issue is that the LANL and OpTC datasets consist of low-level events, not graphs. As we show in the next section, there are in fact many ways to turn these events into static or temporal graphs with edges labeled as normal or anomalous.
3
Diversity in Dataset Preprocessing and Labeling
In order to highlight the heterogeneity of evaluation settings found in the literature, we now study the preprocessing and labeling practices of several published lateral movement detection papers. Note that these practices are not always accurately described in the papers themselves; some of the reported information
6
C. Larroche Table 1. Preprocessing choices for the LANL dataset. Detector
Train/test split
GL-GV [6]
Train: 40 days without red team activity Test: 18 remaining days
Filtering steps
Merging
Computer accounts removed from test set
Train: one edge per source/dest. pair Test: no merging
Only compromised users
No merging
Pikachu [23]
Train: first 40 hours Test: rest of the dataset
Computer and built-in accounts removed Subsampling of users Remote LogOn events only
No merging
Euler [17]
Train: first 41 hours Test: rest of the dataset
Remote NTLM LogOn events only
One edge per snapshot and source/dest. pair
Argus [30]
Train: first 41 hours Test: rest of first 14 days
Remote NTLM LogOn events only
One edge per snapshot and source/dest. pair
Jbeil [14]
Train: first 85% of the NTLM LogOn events dataset for a subset of nodes only Test: last 15% of the dataset
Train: first 12 days
MLTracer [19] Test: next 18 days
No merging
was inferred from the code released by the authors. Sections 3.1 and 3.2 focus on the LANL and OpTC datasets, respectively. 3.1
LANL Dataset
The various preprocessing steps taken in several papers featuring evaluations on the LANL dataset are summarized in Table 1. The first difference between papers lies in the split between training and testing datasets: while Pikachu [23], Euler [17], and Argus [30] stop training right before the first malicious event, MLTracer [19] uses the first 12 days for training, Jbeil [14] trains on the first 85% of the dataset2 , and GL-GV [6] does not enforce a strict temporal separation between the training and testing datasets. MLTracer and Argus also entirely discard the last 28 and 44 days, respectively. As a consequence, the performance metrics originally reported for each of the six detectors were computed over a distinct portion of the dataset. Various filtering steps are also applied: Euler, Argus, and Jbeil only extract NTLM authentications, while GL-GV and Pikachu filter out events related to some users, such as local accounts, computer accounts, and builtin accounts. In addition, Pikachu removes events related to a random subset of user accounts that were not compromised by the red team, and MLTracer only considers events involving compromised accounts. Finally, authentication events mapping to the same (static or temporal) edge are handled in different ways when building the test set. Euler and Argus merge such events 2
This roughly equals 49 days, which appears surprising as no malicious events happen after day 30. A replication study [5] concluded that Jbeil was probably not evaluated on the original LANL labels.
Performance Evaluations for Graph-Based Lateral Movement Detectors
7
Table 2. Preprocessing and labeling choices for the OpTC dataset. Detector
Train/test split Nodes
Edges
Labeling
Internal and external IP addresses (621)
flow start events (40.56M)
All flows from compromised hosts (21,641)
Euler [17]
Train: first 4 days Test: rest of the dataset
Internal IP addresses (1,114) Merge addresses attributed to same host Exclude broadcast/multicast
flow start events with source and All flows from destination addresses compromised mapped to known hosts (21,872) hosts (7.773M)
Argus [30]
Train: first 5 days Test: rest of the dataset
Internal and external IP addresses (814)
flow start events (10.1M)
Train: first 3 days Pikachu [23] Test: rest of the dataset
Unspecified (21,731)
into a single record in the test set, while the other studied papers do not perform such merging. These diverse splitting, filtering, and merging choices lead to entirely different evaluation settings with a direct impact on performance metrics, as further discussed in Section 4.1.
3.2
OpTC Dataset
Table 2 summarizes the preprocessing and labeling choices made by the authors of Pikachu [23], Euler [17], and Argus [30] for the OpTC dataset. Similarly to the LANL dataset, each of them splits the data between training and testing in a different way. However, the most striking difference between them lies in the set of nodes used to build the graphs: while Pikachu and Argus include both internal and external IP addresses, Euler only considers internal ones. In addition, each internal host in the OpTC dataset has several IP addresses; only Euler appears to map these addresses to hostnames so that each host is represented by a single node in the graph. Because of these different preprocessing choices, these three papers report different node counts for the OpTC dataset. Note that these numbers also suggest additional, undocumented preprocessing steps, as Euler reports a greater number of nodes than Pikachu and Argus despite excluding external hosts and merging IP addresses mapping to the same host. Besides leading to varying node counts, these different choices also impact the overall number of flows, as ignoring certain hosts or IP addresses also means dropping all flows involving them. The number of malicious flows varies as well: despite apparently using similar labeling strategies (see Section 5.3 for more details), the three studied papers report different counts. All these differences in the final evaluation dataset can be expected to have a non-negligible impact on the reported performance metrics.
8
4
C. Larroche
Apples and Oranges: Preprocessing the Test Set
The previous section highlighted differences in the evaluation datasets built from the raw LANL and OpTC events by various authors. We explain how these inconsistencies impact the fairness and realism of performance evaluations in Section 4.1, then provide recommendations regarding test set construction for the LANL dataset in Section 4.2 and for the OpTC dataset in Section 4.3. 4.1
Impact of Inconsistent Definitions on Fairness and Realism
As shown in the previous section, preprocessing methods found in the literature differ along three main dimensions: which events are included, how to split the data into training and testing sets, and how to merge events into graph edges. The choices made within each of these categories have a significant impact on the relevance and accuracy of all downstream results. In particular, filtering steps applied to the raw events can lead to unrealistic evaluation settings and artificially inflated detection performance. This is especially clear for the downsampling of benign users described in Section 3.1: removing from the test set all data related to users that are not involved in malicious activity mechanically reduces the number of false positives and thus improves performance metrics. However, in practice, defenders do not know a priori which user accounts are compromised and thus cannot implement this measure. The same observation applies to other filtering rules that more or less subtly leverage specific knowledge about the events of interest: unless they are justified by more general considerations (e.g., the fact that lateral movement only happens between internal hosts), these rules would not exist in a realworld deployment of the evaluated detector. Empirical results obtained in such unrealistic settings are of limited relevance for cybersecurity practitioners. Inconsistent preprocessing choices also negatively impact the fairness of comparisons between detectors, as each detector is evaluated on a different test set. This is most obvious when different train/test splits are made, but even with the same split, the actual set of events for which performance metrics are computed can vary a lot. In particular, different event filtering and merging policies lead to test sets of varying size and class imbalance, both of which directly relate to the difficulty of the detection task. Directly comparing performance metrics reported in different papers therefore leads to biased conclusions, forcing authors of new detection methods to re-evaluate all competing methods on their own test set. This defeats the purpose of benchmark datasets, which are supposed to accelerate research by providing a standardized evaluation procedure. In practice, biased comparisons between performance metrics computed on different test sets are commonly found in the literature [17,5,27]. To illustrate the impact of dataset preprocessing on the fairness of performance evaluations, consider discrete-time graphs with different snapshot lengths. The authors of Euler [17] and Argus [30] investigated the influence of snapshot length on detection performance for the LANL and OpTC datasets, and both found a significant impact. However, they computed performance metrics at the
time
Performance Evaluations for Graph-Based Lateral Movement Detectors A
B
A
C
B
C
A
C
B
E
A
D
B
E
A
C
B
E
9
Δ1: 9 edges total, 3 malicious
Δ1 Δ2
A
t1
B C
B
C
t2
A
A
D
A
B
E
B
t3
C
t4
E
E
B
t5
Δ2: 6 edges total, 3 malicious
A
t1
E
C
C B
A
B D
t2
Fig. 2. Influence of snapshot length on the statistical properties of discrete-time graphs: for the same event sequence, different snapshot lengths ∆1 and ∆2 result in edge sets of different size and class imbalance.
edge level; that is, after merging all events with the same source, destination, and snapshot index into a single record. As a consequence, the number of benign and malicious test samples depends on snapshot length (see Figure 2 for an illustration). Is the change in detection performance due to an actual variation in the accuracy of the detector or to the differences in the composition of the test set for various snapshot lengths? In order to distinguish these two causes, we reproduced the original evaluations of Euler and Argus for different snapshot lengths; for each snapshot length, we then computed the AUC and AP both at the edge level and at the event level (that is, accounting for all the events making up a single edge). The results are shown in Figure 3. There are notable differences in the evolution of some metrics for increasing snapshot lengths depending on the level at which they are computed: in particular, the AP follows different trends for both detectors and both datasets. Diverging trends are also observed for the AUC of Argus on the LANL dataset. These results confirm that event merging affects edge-level performance metrics not only by making detectors more or less accurate, but also directly through the statistical properties of the test set. Consequently, fair comparisons require that performance metrics be computed at the event level, with the exact same events in the test set for all competing detectors. 4.2
Recommendations for the LANL Dataset
The main source of disparity in evaluations on the LANL dataset is the exclusion of some user accounts or authentication packages. Our main recommendation is thus to include all remote LogOn events from the testing period into the test set. More specifically, excluding authentications that do not use NTLM has no justification as lateral movement can also rely on other protocols (such as Kerberos). Because lateral movements in the LANL dataset happen to leverage NTLM, filtering out other authentication packages makes detection easier by reducing class imbalance, thereby artificially inflating performance metrics. The
C. Larroche
15
99.4
10
99.2
5
99.0
Euler/LANL 97.5 96.5
0.2
96.0
0.1
95.5
70
98
60
97
50
96
40
95
30
Euler/OpTC 70
99.5
0.3
97.0
AP (duplicates)
Argus/OpTC
99
0.4
AP (%)
AUC (%)
20
AP (%)
AUC (%)
99.6
AP (original)
AP (%)
AUC (duplicates)
AUC (%)
99.8
Argus/LANL
60
99.0
50
98.5
40
AP (%)
AUC (original)
AUC (%)
10
30
98.0 360 900 1800 3600 Snapshot length (seconds)
360 900 1800 3600 Snapshot length (seconds)
Fig. 3. Performance of Argus and Euler on the LANL and OpTC datasets as a function of snapshot length, with the original preprocessing and after duplication of the merged events. The error bars represent the standard deviation over ten runs. Horizontal jitter added for readability.
same argument can be made about the exclusion of events involving local, builtin, and computer accounts: some of the malicious authentications in the LANL dataset actually do involve local user accounts. Moreover, even though none of them relies on built-in or computer accounts, this is not necessarily true in all attack scenarios. Computer accounts, for instance, can authenticate remotely and be granted all sorts of access rights, which makes them usable for lateral movement. Similarly, some lateral movement tools (such as PsExec) leverage built-in accounts (such as system) to remotely execute commands with elevated privileges. Finally, downsampling or entirely excluding user accounts that are not involved in any malicious event unrealistically decreases the false positive rate and should thus be avoided, as mentioned in Section 4.1. Different ways of splitting the dataset between training and testing also contribute to the heterogeneity of evaluation practices. For the sake of realism, it is preferable to avoid any temporal overlap between the training and testing datasets since a real-world detector can only be trained on past data. It is also sensible to stop training before the first malicious actions take place. Finally, even though the red team exercise stops on day 30, including the remaining 28 days into the evaluation dataset allows for a more accurate estimation of the false positive rate. We thus recommend training on the first 41 hours and testing on the rest of the dataset.
Performance Evaluations for Graph-Based Lateral Movement Detectors
4.3
11
Recommendations for the OpTC Dataset
Using the OpTC dataset to evaluate lateral movement detectors is difficult due to its size, complexity, incompleteness, and lack of documentation. In particular, because no system logs are available for half of the workstations and all of the servers, some of the internal network flows cannot be observed. Besides, some IP addresses cannot be attributed to a specific computer. These difficulties likely explain the diversity of preprocessing methods observed in Section 3.2. However, sensible guidelines can still be provided. First of all, internal network flows should be extracted from the flow start system events, since the network logs only record flows between internal and external hosts. Because lateral movements happen between internal hosts, flows involving an external IP address should be dropped. It is also reasonable to filter out broadcast and multicast flows: though it is technically possible to use broadcast or multicast traffic for lateral movement, well-known techniques observed in the wild rely on unicast traffic. When system logs are available for both the source and destination of a flow, two distinct flow start events are observed (one at each end of the connection). Such duplicates should be removed so that each flow is accounted for only once. This can be done by matching flow start events with the same source and destination addresses and ports and the same transport protocol. In order to make graph-based detectors more accurate, it is helpful to represent all IP addresses attributed to the same host as a single node (the specific address used for a given flow can still be treated as a feature of the corresponding edge). Each host in the simulated enterprise network has three network interfaces, with IPv4 addresses within subnets 142.20.56.0/21, 10.20.0.0/16, and 10.50.0.0/16, respectively. There is also one link-local IPv6 address for each of these interfaces, totaling six addresses per host. It is especially important to include flows related to link-local IPv6 addresses into the test set as some of them result from lateral movements. IPv4 addresses can be attributed to hosts based on the predictable addressing schemes of the three subnets. For link-local IPv6 addresses, information can be extracted from flow start events: since the event is recorded on either the source or destination host, the source (resp. destination) address of an outbound (resp. inbound) flow can be tied to the name of the host recording the event. However, hosts for which no system logs are available cannot be easily matched with their link-local addresses. Finally, the OpTC dataset has a built-in train/test split: the first six days (17-22 September) contain only benign activity, while the last three days (23-25 September) represent one attack scenario each. We thus recommend using the last three days for testing.
5
Moving Targets: Labeling Lateral Movements
Having extracted an evaluation dataset from a given collection of events, one must still decide which elements of this test set should be flagged as lateral
12
C. Larroche
movements. We explain in Section 5.1 why this is not as straightforward as it may appear. Guidelines specific to the LANL and OpTC datasets are then provided in Section 5.2 and Section 5.3, respectively. 5.1
Challenges Related to Labeling
Two main challenges stand in the way of determining whether one specific authentication event or network flow should be flagged as a lateral movement. First, precisely defining what counts as a lateral movement is not straightforward: must there be malicious follow-up activity on the target host? Should failed attempts be taken into account? What about adjacent but distinct tactics, such as network reconnaissance? Even though, strictly speaking, such actions are not lateral movements, detecting them is still valuable for defenders. Second, there is no definitive way to decide exactly which events should be labeled as malicious: only those recorded at the exact moment when a lateral movement happened, or rather all events that would not have been there in the absence of malicious activity? Different labeling choices obviously lead to different performance evaluations. It is thus especially important that these choices be well-grounded and shared across the research community. This is however not yet the case, especially for the OpTC dataset, for which no labels were provided upon release. We thus study the labeling strategies found in the literature and provide guidelines for the LANL and OpTC datasets. 5.2
Recommendations for the LANL Dataset
The LANL dataset features some labeled malicious authentication events. The official description of the dataset [13] does not explicitly qualify these events as lateral movements and warns that they represent only a subset of all events related to red team activity. However, the relatively low granularity of the data makes it difficult to reliably identify additional malicious events or infer the high-level tactic (lateral movement or otherwise) associated with each labeled event. Therefore, using the original labeling to evaluate lateral movement detectors is the most sensible approach despite the aforementioned limitations. Note that all labeled malicious authentications in the LANL dataset are part of one single (simulated) intrusion; in fact, almost all of them come from the same source host (C17693). This has two major implications. First, detecting half of these events does not mean that the detector would only detect half of all intrusions: since most intrusions comprise several lateral movements, detecting only some of these can be considered sufficient; further investigation by a security analyst can then uncover the rest. Second, using part of the labeled events for supervised learning (as done, e.g., in [19]) leads to unrealistic detection performance: in a real-world setting, defenders are not interested in learning from the first half of a single intrusion to detect the second half. The LANL dataset should thus only be used to evaluate unsupervised learning methods (or detectors trained in a supervised manner on another dataset).
Performance Evaluations for Graph-Based Lateral Movement Detectors
13
No system logs svchost.exe
IPC (not tracked)
DC1 Found by taint propagation
SysClient0501
Network flow
Missed by taint propagation
SysClient0010
powershell.exe (malicious)
SysClient0402
svchost.exe
IPC (not tracked) wmiprvse.exe
Spawns powershell.exe (malicious)
SysClient0660
System logs available
Fig. 4. Two examples of lateral movements missed by process-based labeling: (left) no system logs are available for the source host; (right) the process handling the network connection is not the malicious process. Both happen on 24 September.
5.3
Recommendations for the OpTC Dataset
In constrast with the LANL dataset, the OpTC dataset contains no explicit, event-level labels. A high-level description of the actions taken by the red team as well as a few technical indicators (host names, IP addresses, process IDs) are provided instead. This gives researchers more latitude in deciding which events should be treated as malicious. The most common approach in lateral movement detection research has so far consisted in labeling all network flows coming out of a compromised host as lateral movements [23,17]. While that policy is indeed likely to identify all lateral movements, it also falsely labels all legitimate traffic coming from the compromised hosts as malicious. In addition, it labels other kinds of malicious activity (e.g., reconnaissance) as lateral movement. This overly extensive definition introduces bias in performance evaluations: it reduces class imbalance by producing many positive samples, and it counts false positives (resp. true negatives) as true positives (resp. false negatives) by erroneously labeling legitimate traffic as malicious. A more accurate and realistic labeling strategy relying on taint propagation was proposed in previous work studying the OpTC dataset [22,21] (although not for the specific purpose of evaluating lateral movement detectors). Since the PIDs of the main malicious processes are known, it is possible to label events (including flow start events) involving these processes and their children as malicious. This is more specific than the aforementioned hostbased approach but still does not distinguish actual lateral movements from other kinds of malicious traffic. In addition, process-based labeling does not in fact identify all malicious flow start events. This happens for two reasons, as illustrated in Figure 4: first, system logs are unavailable for some of the compromised hosts. As a consequence, malicious flows from these hosts to targets for which system logs are available are recorded only on their destination and thus cannot be traced back to known malicious processes. Second, many lateral movement flow events are generated by processes other than malicious processes and their children. WMI, for instance, relies on a built-in service, so
14
C. Larroche
Table 3. Number of malicious network flows found in the OpTC dataset by different labeling methods. Method Host-based [23,17] Process-based [21] Ours
Number of malicious flows ≈25,000 563 (15 LM-related) 1359 (249 LM-related)
malicious flows are attributed to a service process (svchost.exe) that is not spawned by a malicious process. We confirmed manually that some flow start events clearly related to documented red team activity were absent from the set of malicious events provided by Majorczyk et al [21]. To overcome the limitations of process-based labeling, we propose a hybrid labeling approach. First, we extract all flow start events labeled as malicious by Majorczyk et al. and determine which ones can be traced back to documented lateral movements performed by the red team. Those that cannot are marked as malicious but not lateral movement-related. In other words, we rely on the red team’s report to distinguish lateral movements from other malicious activities. Second, we manually inspect the system logs of the compromised hosts to identify additional flow start events that can be reliably attributed to lateral movements and other malicious actions but are not generated by processes identified as malicious. These events are labeled accordingly. As shown in Table 3, we find fewer malicious flow start events than host-based labeling but many more than process-based labeling. This confirms that manual labeling is required to avoid falsely labeling benign flows as malicious and vice versa. In addition, distinguishing lateral movement-related events from events caused by other malicious actions allows us to evaluate detectors on both their ability to detect lateral movements strictly speaking and their more general effectiveness in flagging malicious activity (see Section 6.2 for more details).
6
A Fair and Realistic Evaluation
In the previous sections, we have identified bad evaluation practices in the lateral movement detection literature and proposed better alternatives in terms of fairness and realism. A final question naturally follows: what is the impact of this change in evaluation methodology for existing lateral movement detectors? To answer it, we focus on three published detectors: Pikachu [23], Euler [17], and Argus [30]. The reason for this choice is twofold: first, the corresponding papers are among the most cited in the lateral movement detection literature. Second, all three methods were already evaluated by their authors on both the LANL and OpTC datasets, providing a baseline with which our new results can be compared. More specifically, we focus on the GCN-GRU link detection variant of Euler and the MPNN-GRU variant of Argus. The authors’ implementations [15,29]
Performance Evaluations for Graph-Based Lateral Movement Detectors
15
Table 4. Description of the datasets obtained with our preprocessing and labeling methodology. There are fewer malicious events in OpTC than in Table 3 because of flow deduplication (described in Section 4.3). Dataset
Nodes
Events
Malicious events
LANL OpTC
16,159 897
369,569,626 36,227,823
702 1354 (244 LM-related)
Table 5. Results of our evaluation and originally reported metrics for LANL. Method Pikachu [23] Euler [17] Argus [30]
AUC (new)
AUC (rep.)
AP (new)
AP (rep.)
78.28±0.43 97.97±0.17 98.44±0.78
94±0.4 99.12 99.83
0.00±0.00 0.02±0.00 0.09±0.04
5.23 32.27
are used for these two detectors, with minor changes in order to apply our dataset preprocessing recommendations. As for Pikachu, we use the improved implementation of Bouček and Husák [5,4], again with minor modifications. For each detector, we set hyperparameters to their best-performing values as reported in the original papers. We compute the average and standard deviation over ten runs for each metric to account for the randomness in model training. These results are compared with those reported in the original papers; note that standard deviations were not included in the Euler and Argus papers, and the AP was not reported in the Pikachu paper. We implement the preprocessing recommendations provided in Section 4 for both datasets; Table 4 contains descriptive statistics for the preprocessed datasets. Our results for the LANL (resp. OpTC) dataset are presented in Section 6.1 (resp. Section 6.2). 6.1
Results for the LANL Dataset
The results of our experiments on the LANL dataset are shown in Table 5 alongside the originally reported performance metrics for each evaluated method. The most obvious outcome is that all detectors perform worse in our evaluation setting than in the original experiments. The drop in performance is especially steep for Pikachu in terms of AUC and for Euler and Argus in terms of AP. However, the ranking remains the same, with Argus coming in first and Pikachu last. The relative performance gaps do change; in particular, Argus outperforms Euler by a smaller margin than originally reported. Overall, these observations suggest that the diverse preprocessing choices made for the LANL dataset mostly impact the realism of the evaluation setting: arbitrarily excluding some benign events from the test set leads to an easier evaluation and thus artificially inflated performance metrics. In fact, the greatest drop in AUC is observed for Pikachu, which was originally evaluated under the most unrealistic conditions by dropping all events involving most benign users. In comparison, the effect on fairness remains limited.
16
C. Larroche
Table 6. Results of our evaluation and originally reported metrics for OpTC. "LM only" (resp. "all") means that the metric was computed strictly for lateral movement events (resp. for all malicious events); "rep." stands for "reported". Method Pikachu [23] Euler [17] Argus [30]
6.2
AUC (LM only)
AUC (all)
AUC AP (rep.) (LM only)
50.74±0.81 86.02±0.68 99±0.03 52.90±6.82 84.31±3.55 88.8 43.23±3.05 79.75±2.19 99.7
AP AP (all) (rep.)
0.01±0.00 0.52±0.06 0.00±0.00 0.51±0.32 0.01±0.00 3.36±1.10
8.8 80.74
Results for the OpTC Dataset
As explained in Section 5.3, we defined two sets of labels for the OpTC dataset: the first one flags only events related to lateral movements strictly speaking, while the second one features all malicious events. We thus compute performance metrics for both labeling policies. The results are displayed in Table 6. Similarly to the LANL dataset, detection performance is worse than reported for all detectors, metrics, and labeling policies. Moreover, all three detectors perform especially badly when trying to detect lateral movements only. This is unsurprising as other kinds of malicious activity, such as network scanning, are more noisy and thus easier to detect. Still, it does confirm the importance of strict labeling in assessing the ability to detect actual lateral movements. Finally, in addition to absolute performance, the ranking of the three detectors is also impacted: Argus, which reportedly largely outperformed competitors, does not do any better than them in our evaluation setting; in contrast, Euler goes from distant last to tied for first in terms of AUC. We thus conclude that previous preprocessing and labeling methodologies for the OpTC dataset significantly impair both fairness and realism.
7
Conclusion
Dataset preprocessing and labeling tend to be overlooked and insufficiently documented when reporting evaluations of lateral movement detectors, leading to diverse and sometimes questionable practices. We studied both this diversity and its impact on the fairness and realism of performance evaluations for the popular LANL and OpTC datasets, empirically showing that previously published lateral movement detection results were excessively optimistic. To encourage more rigorous evaluations, we provide detailed guidelines for properly preprocessing and labeling both datasets. The poor performance of studied detectors in our evaluation setting shows that lateral movement detection remains an open problem. New contributions should be reviewed with this in mind: expecting production-ready performance from research prototypes incentivizes biased rather than honest and informative evaluations. We hope that our guidelines will contribute to moving the field forward through more carefully assessed improvements.
Performance Evaluations for Graph-Based Lateral Movement Detectors
17
References 1. Lateral movement, tactic TA0008. https://attack.mitre.org/tactics/TA0008/, accessed: 2025-03-31 2. Apruzzese, G., Pierazzi, F., Colajanni, M., Marchetti, M.: Detection and threat prioritization of pivoting attacks in large networks. IEEE Trans. Emerg. Top. Comput. 8(2), 404–415 (2020) 3. Arp, D., Quiring, E., Pendlebury, F., Warnecke, A., Pierazzi, F., Wressnegger, C., Cavallaro, L., Rieck, K.: Dos and don’ts of machine learning in computer security. In: USENIX Security (2022) 4. Bouček, V.: Pikachu, https://github.com/vboucek/Pikachu, accessed: 2026-0520 5. Bouček, V., Husák, M.: Replication: Network-based lateral movement detection methods using machine learning. In: CNSM (2025) 6. Bowman, B., Laprade, C., Ji, Y., Huang, H.H.: Detecting lateral movement in enterprise computer networks with unsupervised graph AI. In: RAID (2020) 7. Corso, G., Stark, H., Jegelka, S., Jaakkola, T., Barzilay, R.: Graph neural networks. Nat. Rev. Methods Primers 4(1), 17 (2024) 8. DARPA: Operationally transparent cyber dataset (2020), https://github.com/ FiveDirections/OpTC-data, accessed: 2026-05-20 9. Grover, A., Leskovec, J.: node2vec: Scalable feature learning for networks. In: KDD (2016) 10. GWU GraphLab: Picodomain (2020), https://github.com/iHeartGraph/ PicoDomain, accessed: 2026-05-20 11. Ho, G., Dhiman, M., Akhawe, D., Paxson, V., Savage, S., Voelker, G.M., Wagner, D.: Hopper: Modeling and detecting lateral movement. In: USENIX Security (2021) 12. Kent, A.D.: Comprehensive, multi-source cyber-security events. Los Alamos National Laboratory (2015). https://doi.org/10.17021/1179829 13. Kent, A.D.: Cybersecurity data sources for dynamic network research. In: Dynamic Networks in Cybersecurity. Imperial College Press (2015) 14. Khoury, J., Klisura, Ð., Zanddizari, H., Parra, G.D.L.T., Najafirad, P., Bou-Harb, E.: Jbeil: Temporal graph-based inductive learning to infer lateral movement in evolving enterprise networks. In: SP (2024) 15. King, I.J.: Euler, https://github.com/iHeartGraph/Euler, accessed: 2026-05-20 16. King, I.J., Huang, H.H.: EULER: Detecting network lateral movement via scalable temporal link prediction. In: NDSS (2022) 17. King, I.J., Huang, H.H.: Euler: Detecting network lateral movement via scalable temporal link prediction. ACM Trans. Priv. Secur. 26(3), 35:1–35:36 (2023) 18. Lee, W., McCormick, T.H., Neil, J., Sodja, C., Cui, Y.: Anomaly detection in largescale networks with latent space models. Technometrics 64(2), 241–252 (2022) 19. Liu, F., Wen, Y., Wu, Y., Liang, S., Jiang, X., Meng, D.: MLTracer: Malicious logins detection system via graph neural network. In: TrustCom (2020) 20. Liu, Q., Stokes, J.W., Mead, R., Burrell, T., Hellen, I., Lambert, J., Marochko, A., Cui, W.: Latte: Large-scale lateral movement detection. In: MILCOM (2018) 21. Majorczyk, F., Pilastre, B., Dijoud, F.: A new hope for DARPA OpTC. In: ACSAC Workshops (2025) 22. Nikulshin, V., Talhi, C.: Effective IDS under constraints of modern enterprise networks: revisiting the OpTC dataset. In: CIoT (2024) 23. Paudel, R., Huang, H.H.: Pikachu: Temporal walk based dynamic graph embedding for network anomaly detection. In: NOMS (2022)
18
C. Larroche
24. Rossi, E., Chamberlain, B., Frasca, F., Eynard, D., Monti, F., Bronstein, M.: Temporal graph networks for deep learning on dynamic graphs. arXiv preprint arXiv:2006.10637 (2020) 25. Sun, X., Yang, J.: HetGLM: Lateral movement detection by discovering anomalous links with heterogeneous graph neural network. In: IPCCC (2022) 26. Tang, M., Ge, B.: Social-Hunter: A social heuristics-based approach to early unveiling unknown malicious logins using valid accounts. Comput. Secur. p. 104269 (2024) 27. Wang, C., Zheng, P., Gui, J., Hua, C., Hassan, W.U.: R+R: from claims to crashes: A systematic re-evaluation of graph-based network intrusion detection systems. In: ACSAC (2025) 28. Wei, R., Cai, L., Yu, A., Meng, D.: AGE: Authentication graph embedding for detecting anomalous login activities. In: ICICS (2019) 29. Xu, J.: Argus, https://github.com/C0ldstudy/Argus, accessed: 2026-05-20 30. Xu, J., Shu, X., Li, Z.: Understanding and bridging the gap between unsupervised network representation learning and security analytics. In: SP (2024) 31. Zhao, S., Wei, R., Cai, L., Yu, A., Meng, D.: CTLMD: Continuous-temporal lateral movement detection using graph embedding. In: ICICS (2019)