Evaluating Explainable AI for Deep Learning-Based Network Intrusion

Detection System Alert Classiﬁcation

Rajesh Kalakoti

1 a

, Risto Vaarandi

1 b

, Hayretdin Bahs¸i

1,2 c

and Sven N

omm

1 d

Department of Software Science, Tallinn University of Technology, Tallinn, Estonia

School of Informatics, Computing, and Cyber Systems, Northern Arizona University, U.S.A.

Keywords:

Network Intrusion Detection System, NIDS Alerts, SOC, Evaluation of Explainability.

Abstract:

A Network Intrusion Detection System (NIDS) monitors networks for cyber attacks and other unwanted activ-

ities. However, NIDS solutions often generate an overwhelming number of alerts daily, making it challenging

for analysts to prioritize high-priority threats. While deep learning models promise to automate the prioriti-

zation of NIDS alerts, the lack of transparency in these models can undermine trust in their decision-making.

This study highlights the critical need for explainable artiﬁcial intelligence (XAI) in NIDS alert classiﬁcation

to improve trust and interpretability. We employed a real-world NIDS alert dataset from Security Opera-

tions Center (SOC) of TalTech (Tallinn University Of Technology) in Estonia, developing a Long Short-Term

Memory (LSTM) model to prioritize alerts. To explain the LSTM model’s alert prioritization decisions, we

implemented and compared four XAI methods: Local Interpretable Model-Agnostic Explanations (LIME),

SHapley Additive exPlanations (SHAP), Integrated Gradients, and DeepLIFT. The quality of these XAI meth-

ods was assessed using a comprehensive framework that evaluated faithfulness, complexity, robustness, and

reliability. Our results demonstrate that DeepLIFT consistently outperformed the other XAI methods, pro-

viding explanations with high faithfulness, low complexity, robust performance, and strong reliability. In

collaboration with SOC analysts, we identiﬁed key features essential for effective alert classiﬁcation. The

strong alignment between these analyst-identiﬁed features and those obtained by the XAI methods validates

their effectiveness and enhances the practical applicability of our approach.

1 INTRODUCTION

Many organizations use open-source (e.g., Suricata

and Snort) or commercial (e.g., Cisco NGIPS) NIDS

platforms to identify malicious network trafﬁc (Day

and Burns, 2011). Most widely used NIDS platforms

use human-created signatures to identify malicious

network trafﬁc. However, this often results in many

alerts, with only a tiny fraction deserving closer at-

tention from security analysts (Jyothsna et al., 2011).

In a typical SOC operation, security analysts an-

alyze the alerts based on their impact on the secu-

rity of the organizational assets and categorize them

as high or low priority. At this stage, analysts also

identify the false positives that are benign system ac-

tivities but are ﬂagged as alerts by NIDS. Security

https://orcid.org/0000-0001-7390-8034

https://orcid.org/0000-0001-7781-5863

https://orcid.org/0000-0001-8882-4095

https://orcid.org/0000-0001-5571-1692

analysts ﬁnd it challenging to identify high-priority

alerts (Jyothsna et al., 2011). Machine learning (ML)

Deep Learning (DL) methods constitute a signiﬁ-

cant solution to automatize these prioritization tasks

and, thus, reduce SOC workloads, especially in the

lower-tier levels of security monitoring and incident

handling processes in the related literature, with ap-

proaches divided into supervised, unsupervised, and

semi-automated methods (Vaarandi, 2021; Vaarandi

and M

ases, 2022; Kalakoti et al., 2022). However, the

explainability or interpretability of ML models arises

as a signiﬁcant concern in alert prioritization despite

their signiﬁcant contribution.

Explainable Artiﬁcial Intelligence (XAI or Ex-

plainable AI) is necessary for experts to verify alert

classiﬁcations and for industries to comply with reg-

ulations (Goodman and Flaxman, 2017). In cyberse-

curity, it’s vital to explain ﬂagged network activities

as potential threats. XAI helps meet compliance stan-

dards and improve systems by clarifying NIDS alert

classiﬁcations and identifying crucial features for data

Kalakoti, R., Vaarandi, R., Bah¸si, H. and Nõmm, S.

Evaluating Explainable AI for Deep Learning-Based Network Intrusion Detection System Alert Classiﬁcation.

DOI: 10.5220/0013180700003899

In Proceedings of the 11th International Conference on Information Systems Security and Privacy (ICISSP 2025) - Volume 1, pages 47-58

ISBN: 978-989-758-735-1; ISSN: 2184-4356

collection. In the event of a security breach, XAI of-

fers valuable insights for forensic analysis, helping to

understand why speciﬁc alerts were or were not trig-

gered, which is crucial in reconstructing the timeline

and nature of an attack (Alam and Altiparmak, 2024).

NIDS usually struggles with high false positive rates.

XAI can enable security analysts to understand why

particular benign activities are mistakenly ﬂagged as

threats, enabling more transparent system tuning and

reducing false alarms (Moustafa et al., 2023).

Explainable AI (XAI) methods address the model

opacity problem through various global and local ex-

planation methods (Rawal et al., 2021). Several stud-

ies have studied explainable AI methods in intrusion

detection (Alam and Altiparmak, 2024; Szczepa

nski

et al., 2020; Senevirathna et al., 2024; Moustafa et al.,

2023). However, it is crucial to note that these stud-

ies did not comprehensively evaluate Explainable AI

methods under various intrusion datasets and miscel-

laneous sets of Black box nature of AI models. This

lack of comprehensive evaluation signiﬁcantly affects

the generality of such methods, highlighting the ur-

gent need for further research in this area. Although

XAI-based IDS tools are expected to be an integral

part of network security to help security analysts in

SOCs to enhance the efﬁciency and precise in net-

work defence and threat mitigation, a key challenge of

deploying XAI-Based model into network intrusion

detection is assessing such tools, testing their quality,

and evaluating the relevant security metrics. These

challenges undermine the trust in using the XAI-IDS

model for real-world deployment in network IDS sys-

tems.

In this paper, we propose a Long Short-Term

Memory (LSTM) model for NIDS alert prioritiza-

tion to improve transparency and Reliability. This

study evaluates various XAI methods to bridge the

gap between the high accuracy of complex ML mod-

els and the need for transparent, explainable decision-

making in the cybersecurity problem domain. Ob-

jectives of the study include creating an explainable

LSTM model for NIDS alert classiﬁcation, compar-

ing four advanced XAI methods, evaluating their per-

formance using comprehensive metrics, and validat-

ing XAI-generated explanations based on four crite-

ria: Faithfulness, Complexity, Robustness, and Relia-

bility.

Faithfulness estimates how accurately the expla-

nation reﬂects the model’s behaviour, assuring that

the local explanation represents the model’s decision-

making process. Robustness evaluates the stability of

explanations under small input perturbations, which

is vital for building faith in local explanations. Com-

plexity assesses the simplicity of the explanations, as

more detailed explanations are generally more inter-

pretable and valuable for human understanding. Reli-

ability guarantees that the explanations are consistent

with established knowledge, such as the features iden-

tiﬁed by SOC analysts in this case.

We propose that explainable AI methods can pro-

vide explanations for the decision-making processes

of the LSTM model, prioritizing NIDS alerts and ul-

timately boosting the trust and usefulness of these

systems. This research particularly examined a real-

world dataset of NIDS alerts using LSTM, interpret-

ing the output decisions made by these models and

evaluating them through both quantitative and quali-

tative (expert) evaluations. This study emphasizes ar-

tiﬁcial intelligence (XAI) in high-risk threat detection

settings. Our research offers a perspective to the exist-

ing literature as the aspect of interpretability has not

been explored in relation to the signiﬁcance of NIDS

alerts. This research suggests that a well-designed

benchmarking study can identify high-performance

detection models that provide high-quality explana-

tions. Therefore, security experts may not need to

sacriﬁce detection performance over a model for ex-

plainability in the addressed ML studies.

Our paper is structured as follows: Section 2 re-

views related work on NIDS and XAI in NIDS, Sec-

tion 3 outlines our methodology, Section 4 presents

our results and discussions, and Section 5 provides

our conclusions.

2 RELATED WORK

ML and DL have advanced the analysis of NIDS

alerts. This section reviews key contributions in NIDS

alert processing, focusing on classiﬁcation, cluster-

ing, and explainable AI methods. It delves into stud-

ies addressing challenges such as alert prioritization,

false positive reduction, and interpretable models in

cybersecurity.

(Kidmose et al., 2020) proposed a three-phase

method for NIDS alert classiﬁcation (Kidmose et al.,

2020). They used an LSTM and latent semantic anal-

ysis to convert textual alerts into vectors, clustered

the vectors using the DBSCAN algorithm, and clas-

siﬁed incoming alerts based on their similarity to the

core points of the clusters. (Van Ede et al., 2022)

developed a semi-automated method for classifying

NIDS alerts and other security events,which involved

detecting and analyzing event sequences using deep

learning models, clustering with the DBSCAN algo-

rithm, and human analysts labeling the resulting clus-

ters (Van Ede et al., 2022). Labeled database was then

used for semi-automated classiﬁcation of additional

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy

event sequences, with human analysts manually re-

viewing unclustered events.

In a paper(Mane and Rao, 2021), the authors uti-

lized SHAP, LIME, Contrastive Explanations Method

(CEM), ProtoDash, and Boolean Decision Rules via

Column Generation (BRCG) over the NSL-KDD

dataset (Tavallaee et al., 2009) for intrusion detection

system (IDS). They demonstrated the factors that in-

ﬂuence the prediction of cyber-attacks.

(Ban et al., 2023) proposed a method using

an IWSVM-based classiﬁer to detect critical NIDS

alerts. The classiﬁer assigned higher weights to re-

peated data points and the minority class of critical

alerts. A clustering algorithm grouped alerts repre-

senting the same incident based on attributes such

as IP addresses, service ports, and alert occurrence

time. (Shin et al., 2019) developed an organizational

platform using machine learning to analyze NIDS

alert data with support for binary SVM and one-class

SVM methods (Shin et al., 2019). In a paper (Feng

et al., 2017), authors described another organizational

implementation for processing NIDS alerts and other

security events to identify at-risk users. (Wang et al.,

2019) used a graph-based method to eliminate false

alerts and applied GBDT algorithms for alert classi-

ﬁcation. (Ban et al., 2021) used a large NIDS dataset

to evaluate seven supervised machine learning meth-

ods (Ban et al., 2021). They found that Weighted

SVM, SVM, and AB (Adaboost) produced the best

results, while two isolation forest-based unsupervised

algorithms provided lower precision than the evalu-

ated supervised algorithms.

It is important to note that a large body of research

is devoted to replacing NIDS with ML-based sys-

tems (Tsai et al., 2009). However, organizations use

signature-based NIDSs due to the wide availability of

this technology and complex SOC processes evolving

around these systems. Thus, prioritizing NIDS alerts

is a signiﬁcant real-world challenge in SOCs. Vari-

ous research studies have addressed the explainability

of ML-based NIDS systems. However, to our knowl-

edge, the explainability of the ML models developed

for NIDS alert prioritization has not been studied in

the literature.

(Szczepa

nski et al., 2020) introduced the hybrid

Oracle Explainer IDS, which combines artiﬁcial neu-

ral networks and decision trees to achieve high ac-

curacy and provide human-understandable explana-

tions for its decisions (Szczepa

nski et al., 2020). In

a paper (Senevirathna et al., 2024), authors have de-

veloped an Oracle-based Explainer module that uses

the closest cluster to generate an explanation for the

decision. A study explores how explanations in the

context of 5G security can be targeted and weakened

using scaffolding techniques. The authors suggest

a framework for carrying out the scaffolding attack

within a security setting, which involves selecting fea-

tures and training models by combining explainable

AI methods. (Zolanvari et al., 2021)(Zolanvari et al.,

2021) introduced a model-agnostic XAI framework

called TRUST for numerical applications. It uses fac-

tor analysis to transform input features, mutual infor-

mation to rank features, and a multimodal Gaussian

distribution to generate new samples for each class

label.

Some other studies have explored explainable AI

methods in intrusion detection (Alam and Altiparmak,

2024; Szczepa

nski et al., 2020; Kumar and Thing,

2024; Kalakoti et al., 2024a; Kalakoti et al., 2024c;

Kalakoti et al., 2024b; Kalakoti et al., 2023). In con-

trast to studies on machine learning-based Network

Intrusion Detection Systems (NIDSs), our research

emphasizes the signiﬁcance of making NIDS alerts

understandable through model transparency. Our ap-

proach incorporates eXplainable AI (XAI) techniques

to evaluate their effectiveness in clarifying NIDS alert

classiﬁcations. We worked with a real world NIDS

dataset from an environment making our ﬁndings

more relevant than those based on old data sets. Our

evaluation criteria cover aspects such as the reliabil-

ity, faithfulness, robustness and complexity of expla-

nations assessing explainability within this domain.

By engaging Security Operations Center (SOC) an-

alysts in verifying our XAI ﬁndings we bridge the

gap, between machine learning models and human

knowledge. This progress enhances XAI in the ﬁeld

of cybersecurity, offering perspectives for developing

transparent and reliable NIDS alert critical prioritiza-

tion systems.

3 METHODOLOGY

3.1 Dataset

Our study makes use of a NIDS alert dataset taken

from a Suricata NIDS system deployed at the Secu-

rity Operations Center (SOC) of Tallinn University

of Technology (Taltech). The dataset was gathered

using the Customized Stream Clustering Algorithm

for Suricata (CSCAS) to analyze alerts from Suricata

NIDS at TalTechs SOC. Data was collected over a

span of 60 days, from January to March 2022 dur-

ing which Suricata generated alerts, for network ac-

tivity involving 45,339 hosts and 4401 TalTech hosts.

The categorized dataset can be accessed at the link;

https://github.com/ristov/nids-alert-data.

Throughout the data collection phase CSCAS op-

Evaluating Explainable AI for Deep Learning-Based Network Intrusion Detection System Alert Classiﬁcation

erated with settings; SessionLength = 300 seconds

(5 minutes) SessionTimeout = 60 seconds (1 minute)

ClusterTimeout = 604,800 seconds (1 week) Cand-

Timeout = 36,000 seconds (10 hours) MaxCandAge =

864,000 seconds (10 days) and α = 0.01. These con-

ﬁgurations have been employed for CSCAS in an en-

vironment since 2021 and were determined to be opti-

mal as outlined in (Vaarandi, 2021). NIDS Alerts are

classiﬁed as either ”important” or ”irrelevant.” Data

points of network trafﬁc were generated by a cus-

tomized version of SCAS, a stream clustering algo-

rithm, and have labels indicating whether they are re-

garded as inliers or outliers by SCAS. Data points are

labeled by humans to indicate if they represent im-

portant or irrelevant alert groups. Important alerts are

prioritized in the SOC security monitoring processes.

Irrelevant alerts include low-priority threats (e.g., fre-

quent scanning for old vulnerabilities) or false pos-

itives (e.g., alerts related to attempts to resolve bot-

net C&C server DNS names not originating from in-

fected computers but from speciﬁc security applica-

tions). The description of the dataset (Vaarandi and

Guerra-Manzanares, 2024) is given below:

• Timestamp – alert group reporting time

• SignatureText – human readable alert text

• SignatureID – numerical signature ID

• SignatureMatchesPerDay – Average matches per

day by the triggering signature (set to 0 if ﬁrst

match was less than 24 hours ago).

• AlertCount – the number of alerts in the current

alert group

• Proto – numerical protocol ID (e.g., 6 denotes

TCP and 17 UDP)

• ExtIP – anonymized IP address of the external

host (extipN, where N is a number that identiﬁes

the given IP address)

• ExtPort – port at the external host, set to -1 if

alerts involve multiple external ports

• IntIP – Anonymized IP address of the internal

host (intipN), set to -1 if alerts involved multiple

internal IP addresses.

• IntPort – port at the internal host, set to -1 if

alerts involve multiple internal ports.

• Similarity – The overall similarity of this alert

group to others in the same cluster or, if it’s an

outlier, to other outlier alert groups. The value

ranges from 0 to 1, with higher values indicating

a high degree of similarity.

• SCAS – The label assigned by the customized

version of SCAS. Here, 0 denotes an inlier and 1

denotes an outlier.

• AttrSimilarity – similarity for the network IDS

alert attribute Attr (there are 34 attributes in to-

tal). Set to -1 if the attribute Attr is not set for the

given signature, otherwise ranges from 0 to 1.

The ﬁeld indicates how often the attribute value

has been observed in other alert groups from the

same cluster (or in other outlier alert groups if

the current alert group is an outlier).

We collaborated with Security Operations Center

(SOC) analysts from TalTech, Estonia to estimate the

reliability of the post-hoc explanations generated for

the decisions of the black-box model, which is the

DL model induced for alert classiﬁcation in this work.

A detailed description of TalTech SOC can be found

in (Vaarandi and M

ases, 2022). Leveraging their ex-

pertise in managing Network Intrusion Detection Sys-

tem (NIDS) alerts, the SOC team at TalTech identiﬁed

the ﬁve features for determining alert signiﬁcance as

outlined in Table 1. These features act as benchmark-

ing reference features in our research to evaluate how

well our XAI algorithms perform.

Table 1: Key Features Identiﬁed by Taltech SOC Analyst

for Determining NIDS Alert Signiﬁcance.

SignatureMatchesPerDay

Similarity

SCAS

SignatureID

SignatureIDSimilarity

For our work, the dataset excluded ’Signature-

Text’ and ’Timestamp’ features as external IP ad-

dresses (”ExtIP” feature) and internal IP addresses

(”IntIP” feature) prior, to model training.

3.2 Long Short-Term Memory for NIDS

Alerts

In this study, we proposed long-term memory

(LSTM) to classify whether a given NIDS alert group

needs immediate attention (Important class label) or

can be assessed as less critical (Irrelevant class label).

LSTM is a neural network designed to address the

long-term dependence problem in traditional recur-

rent neural networks. It introduces forget, input, and

output gates to control the ﬂow of information and

maintain long-term memory. Figure 1 shows struc-

ture of the hidden layer of the LSTM network. The

forget gate adapts to the context, discarding unneces-

sary information. It uses a sigmoid function to pro-

duce a value between 0 and 1, then multiplied by the

previous cell state. A value of 0 means complete for-

getting, while 1 means fully retained.

= σ(W

· [h

t−1

, x

] + b

) (1)

The input gate enhances the necessary informa-

tion for the new cell state, and its output is a sigmoid

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy

Figure 1: Hidden Layer Architecture of LSTM Network.

function with a range of 0 to 1, which is multiplied by

the current cell state.

= σ(W

· [h

t−1

, x

] + b

) (2)

= tanh(W

· [h

t−1

, x

] + b

) (3)

Then the old and new state information can be

combined to construct the ﬁnal new cell state.

= f

×C

t−1

+ i

(4)

The output is determined by the output gate,

which uses a sigmoid function to select information to

be output along with the ﬁnal cell state and the Tanh

function.

= σ(W

· [h

t−1

, x

] + b

) (5)

= O

× tanh(C

) (6)

For training LSTM model, We selected 10,000

data points for each class label (’irrelevant’ and ’im-

portant’), resulting in a total of 20,000 samples. The

The dataset was divided into training and testing sets

at an 80 20-split ratio. We applied the data normaliza-

tion technique to the dataset to convert the values to a

standard scale. We used Min-Max normalization, one

of several available techniques, to transform and nor-

malize the input features to scale them within a range

of 0 to 1, as shown in Equation 7.

′

x − x

min

max

− x

min

(7)

where x

min

is the smallest value of the feature, x

max

is the largest value of the feature, and x is the ac-

tual value of the feature. The normalized feature, x

′

ranges between 0 and 1.

We used RandomSearch hyperparameter tuning

with Ray Tune library

to train LSTM model. We

evaluated the performance of LSTM model for NIDS

alerts classiﬁcation using a confusion matrix. In

NIDS alerts classiﬁcation, True Positives (TP) are the

https://docs.ray.io/en/latest/tune/index.html

number of important alerts correctly classiﬁed as im-

portant, True Negatives (TN) are the number of irrele-

vant alerts correctly classiﬁed as irrelevant, False Pos-

itives (FP) are the number of irrelevant alerts incor-

rectly classiﬁed as important. False Negatives (FN)

are the number of important alerts incorrectly classi-

ﬁed as irrelevant. we used the following evaluation

metrics

Accuracy =

T P + T N

T P + T N + FP + FN

(8)

Precision =

T P

T P + FP

(9)

Recall =

T P

T P + FN

(10)

F1-Score = 2 ×

Precision × Recall

Precision + Recall

(11)

We used softmax activation function at the output

layer to predict class labels, which provides predic-

tion probabilities for each class and enables us to un-

derstand the model’s conﬁdence and the probability

distribution. It’s also crucial to evaluate XAI tech-

niques based on metrics like faithfulness, monotonic-

ity and max sensitivity as discussed in section 3.4.

3.3 Explainable AI Methods

When explaining the model using Explainable AI,

there are two approaches: model agnostic and model

speciﬁc. Explainable AI methods are also categorized

into two types explanations. Local explanations in-

terpret individual predictions and global explanations

that offer an overview of the model’s behaviour. Our

goal is to enhance the explainability of NIDS alerts

detected by LSTM model. We have utilized four pop-

ular XAI feature attribution methods. Will provide a

brief overview of each one. The following outlines

the four methods (LIME, SHAP, Integrated Gradients

(IG) and DeepLIFT) all designed to clarify instances

and shed light on how the model makes decisions, for

speciﬁc predictions. Let x ∈ R

be the input, where

d is the feature set dimensionality. The black box

model M maps input to output M (x) ∈ Y . Dataset

D = (x

, y

) contains all input-output pairs. The expla-

nation mapping g for predictor M and point x returns

importance scores g(M , x) = φ

∈ R

for all features.

Let D : R

× R

7→ R≥ 0 be a metric in the expla-

nation space and S : R

× R

7→ R≥ 0 a metric in the

input space. The evaluation criterion µ maps predictor

M , explainer g, and point x to a scalar.

Evaluating Explainable AI for Deep Learning-Based Network Intrusion Detection System Alert Classiﬁcation

3.3.1 SHAP

SHAP (Lundberg and Lee, 2017) uses Shapley val-

ues from game theory to attribute the importance of

each feature to a model’s prediction, providing a uni-

ﬁed measure of feature importance. SHAP based

on Shapley values, is deﬁned as: g(M , x) = φ

∑

j=1

where φ

is the feature attribution of feature

j. SHAP’s DeepExplainer was used in this study.

3.3.2 LIME

LIME (Ribeiro et al., 2016) (Local Interpretable

Model-agnostic Explanations) constructs a locally in-

terpretable model around a speciﬁc prediction. It

works by perturbing the input and ﬁtting a simple

model, like a linear model, to explain the behaviour of

the black box model in the vicinity of the prediction of

interest. LIME approximates model behavior locally

around (x) by minimizing: argmin

g∈G

L(M , g, π

)+ Ω(g)

where g is an interpretable model in the neighborhood

of (x).

3.3.3 Integrated Gradients

Integrated Gradients (IG) (Sundararajan et al., 2017)

attributes the prediction of a deep network to its inputs

by integrating the gradients along a straight-line path

from a baseline input to the actual input. This method

satisﬁes desirable axioms like completeness and sen-

sitivity, providing a theoretically sound approach

to feature attribution. IG attributes feature impor-

tance by integrating model gradients from a baseline

g(M , x) = IG(x) = (x − ¯x) ×

α=0

∂M ( ¯x+α·(x− ¯x))

∂x

dα

where ¯x is the baseline input.

3.3.4 DeepLIFT

DeepLIFT (Shrikumar et al., 2017) assigns each in-

put (x) a value C

∆x

∆y

representing its deviation from

a reference value, satisfying:

∑

i=1

∆x

∆o

= ∆o where

o = M (x) and ∆o is the difference between model

output and reference value.

3.4 Evaluation of Explainable AI

Methods

The evaluation of Explainable AI methods is cru-

cial to ensure that the explanations they provided

are transparent, also accurate and reliable. We em-

ploy four key metrics to assess the quality of our

explanations for LSTM Model based NIDS alerts:

Reliability, Faithfulness, Robustness and Complex-

ity. These metrics provide a comprehensive evalu-

ation framework that addresses different aspects of

explanation quality. XAI evaluation is categorized

into three groups (Coroama and Groza, 2022): user-

focused evaluation, application-focused evaluation,

and functionality-focused evaluation. The ﬁrst two

categories are part of human-centered evaluation and

are broken down into subjective and objective mea-

sures.

3.4.1 Reliability

An explanation should be centered around the region

of interest, the ground truth GT. g(M , x) = GT.

’Major’ parts of an explanation should lie inside the

ground truth mask GT(x) for both Relevance Mass

Accuracy and Relevance Rank Accuracy metrics used

in this work, and the Ground truth mask ([0,1]) was

determined by the features SOC Analysts identiﬁed

(see Table. 1).Truth-based measures relevance rank

accuracy and relevance mask accuracy are derived

from (Arras et al., 2022).

(a) Relevance Rank Accuracy (RRA) (Arras et al.,

2022): Relevance rank accuracy measures how

much of the high-intensity relevance lies within

the ground truth. We sort the top K values of

g(M , x) in decreasing order X

topK

= {x

, ..., x

g(M , x)

> ... > g(M , x)

RRA =

top

∩ GT(x)|

|GT(x)|

Here top

are features Identiﬁed by SOC Ana-

lyst.

(b) Relevance Mass Accuracy (RMA) (Arras et al.,

2022): The relevance mass accuracy is calcu-

lated as the sum of the explanation values within

the ground truth mask divided by the sum of all

values.

RMA =

∑

g(M , x)

· GT(x

)

∑

g(M , x)

3.4.2 Faithfulness

The explanation algorithm g should replicate the

model’s behavior. g(M , x) ≈ M (x). Faithfulness

quantiﬁes the consistency between the prediction

model M and explanation g. For evaluating the

Faithfulness of explanations, the Faithfulness correla-

tion (Bhatt et al., 2020) and Monotonocity (Luss et al.,

2019) metrics were used.

(a) High Faithfulness Correlation: Faithfulness

measures how well the explanation function g

aligns feature importance scores with the black-

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy

box model M

(M , g; x) = corr

B∈

(

|d|

|B|

)

∑

i∈B

g(M , x)

, M (x)−M (x

)

(12)

where x

= x

|i ∈ B} High Faithfulness correla-

tion metric iteratively substitutes a random sub-

set of given attributions with a baseline value

B. Then, it measures the correlation between the

sum of these attributions and the difference in the

model’s output.

(b) Monotonicity: Let x, x

′

∈ R

be two input points

such that x

≤ x

′

for all i ∈ 1, 2, . . . , d. M and g

are said to be monotonic if the following condi-

tion holds: For any subset S ⊆ 1, 2, . . . , d, the sum

of the attributions of the features in S should be

nonnegative when moving from x to x

′

, that is,

∑

i∈S

g(M , x)i ≤

∑

i∈S

g(M , x

′

)

implies

M (x) − M (x

= ¯x

]

) ≤ M (x

′

) − M (x

′

= ¯x

])

3.4.3 Robustness

Robustness refers to similar inputs should result

in similar explanations. g(M , x) ≈ g(M , x +

ε) for small ε.

(a) Max Sensitivity: Max sensitivity (Bhatt et al.,

2020): is used to ensure that nearby inputs with

similar model output have similar explanations,

it is desirable for the explanation function g to

have a low sensitivity in the region surround-

ing the point of interest x, assuming the differ-

entiability of the predictor function M . Maxi-

mum sensitivity of an explanation function g at

a point of interest x in its neighbourhood is de-

ﬁned as follows: Consider a neighbourhood N

of points within a radius r of x, denoted by N

z ∈ D

|p(x, z) ≤ r, M (x) = M (x)(z), where D is

the distance metric, and p is the proximity func-

tion. Given a predictor M (x), a distance metric

D, a proximity function p, a radius r, and a point

x, we deﬁne the maximum sensitivity of g at x as

follows:

(M (x), g, r; x) = max

z∈N

D(g(M (x), x), g(M (x), z))

(13)

3.4.4 Complexity

Explanations using a smaller number of features are

preferred. It is assumed that explanations using a

large number of features are difﬁcult for the user to

understand.min



g(M , x)



(a) Low Complexity: Low complexity (Bhatt et al.,

2020) metric computes the entropy of each fea-

ture’s fractional contribution to the total attribu-

tion magnitude individually.

(M , g; x) = −

∑

i=1

(i)logP

(i) (14)

where

(i) =

|g(M , x)i|

∑

j ∈ |d||g(M , x)

= P

(1), ....P

(d)

(15)

The experiments were carried out on a computer

running Pop! OS 22.04 LTS x86 64 operating system

with the following hardware conﬁguration: 32 GB of

DDR4-2666R ECC RAM, AMD Ryzen 5 5600G with

Radeon Graphics (12) @ 3.900GHz processor. The

scripts were developed using the Python 3.9 program-

ming language and Pytorch library. For the imple-

mentation of the Integrated Gradients and DeepLIFT

explainers, Captum library was used.

4 RESULTS & DISCUSSIONS

In this section, we present the results of our research,

including an analysis of the LSTM model’s perfor-

mance and explanations of LSTM model using Ex-

plainable AI methods and the quality of evaluation for

these explanations based on four criteria: faithfulness,

complexity, reliability, and robustness.

Figure. 3a shows the confusion matrix, indicating

the model’s strong classiﬁcation performance for test

data of 4000. It correctly classiﬁed 2005 important

alerts and 1980 irrelevant alerts, with only 14 misclas-

siﬁcations of irrelevant alerts as necessary, demon-

strating high accuracy and a low false positive rate.

Figure. 2a shows the training and validation loss over

70 epochs obtained through random search parame-

ter tuning. Initially, both decrease rapidly before sta-

bilizing, indicating convergence without overﬁtting.

The close alignment of the training and validation loss

curves represents good generalization to unseen data.

Figure. 2b shows the training and validation accu-

racy, which quickly stabilizes above 99.5%, indicat-

ing strong model performance. In Figure. 3b, from the

classiﬁcation report, the model achieves near-perfect

precision, recall, and F1-score scores for both classes.

In this paper, we utilized 4 different explainable

AI methods (LIME, SHAP, IG, and DeepLift) to ex-

plain the predictions of our LSTM model on the test

data. LIME analyzes how the model assigns prob-

abilities to categories by comparing these probabili-

ties with the actual category of the data point. SHAP

method provides single-data-point explanations for

models, giving insights. In explanations, a particular

Evaluating Explainable AI for Deep Learning-Based Network Intrusion Detection System Alert Classiﬁcation

(a) Loss (b) Accuracy.

Figure 2: Loss and Accuracy from Best LSTM Performance

Model.

(a) Confusion Matrix (b) Classiﬁcation Report

Figure 3: Confusion Matrix and Classiﬁcation Report.

data point is selected to demonstrate how each feature

inﬂuences the model’s prediction.

Fig. 4a shows an local explanation from LIME

method for a NIDS alert labeled as ”Important.”. Left

side presents prediction probabilities with a 100%

probability for the ”Important” class. On the right

side it illustrates the impact of features. For instance,

when the feature ‘SignatureIDSimilarity’ is less than

or equal to 0.01, it positively affects the ”Important”

classiﬁcation of NIDS alert. Additionally, ‘Signa-

tureMatchesPerDay’ and ‘SCAS’ being less than or

equal to 1.00 also contribute positively. Conversely,

‘ExtPortSimilarity’ and ‘TlsSniSimilarity’ have im-

pacts, suggesting that some NIDS alerts may not be

relevant. SHAP employs Shapley values to showcase

how features inﬂuence model predictions in Fig. 4b of

force plot, red bar signiﬁes the positive impact while

blue bar indicates the negative impact on the model

output. Each bar demonstrates whether the features

bring the predicted value closer to or farther from

the base value of 0.02463. The plot’s base value is

the average of all prediction values. Each strip in

the plot displays the impact of the features on mov-

ing the predicted value closer to or farther from the

base value. Final prediction is deemed an ”important

class label”, with a value of 1.00 for this NIDS alert.

Features, like ’IntPort’ (Internal Port) ’SignatureID-

Similarity’. ExtPort’ (External Port) along with ’Sig-

natureID’ play a role in indicating the importance of

NIDS alert. However, the feature ’HttpStatusSimilar-

ity’ might suggest that this alert could be a less critical

feature to its impact.

DeepLift is a technique used to attribute the out-

(a) LIME explanations for important NIDS alerts

using an LSTM model

(b) SHAP explanations for an important NIDS

alert data point using an LSTM model

NIDS alert data point using an LSTM model

(d) Integrated Gradients feature importance for an

important NIDS alert data point using an LSTM

model

Figure 4: Explanations for an important NIDS alert data

point using an LSTM model.

put of LSTM model to its input features by compar-

ing neuron activation to a reference activation and

assigning contribution scores based on the variance.

Fig. 4c illustrates the signiﬁcance of features using the

DeepLift explainer for the 10 features of a NIDS alert

data point labeled as ”important.” The negative attri-

bution of ’SCAS’ suggests its inﬂuence on classifying

as ”Important” in NIDS alerts. Additionally ’Http-

MethodSimilarity’ and ’IntIP’ show negative attribu-

tions while HttpContentTypeSimilarity has a slight

positive impact countering the ”Important” classiﬁca-

tion. IG attribute a LSTM model’s prediction its input

features by integrating gradients of the model’s out-

put with respect to the input along from a baseline to

the input. This explanation technique works best for

models that use linear activation functions. Fig. 4d

showcases feature importance using IG explainer for

a data point in the ”Important” NIDS alert class la-

bel among the 10 features. Features such, as ’Sig-

natureID’ ’SCAS,’ and ’HttpStatusSimilarity’ display

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy

Table 2: Evaluation Results of Explainable AI Methods: Mean (µ) and Standard Deviation (σ) Values.

Explanation Criterion Faithfulness Robustness Complexity Reliability

Explainer/Metric

High

Faithfulnes

Monotonicity Max Sensitvity Low Complexity

Relevance Mass

Accuracy

Relevancy Rank

Accuracy

µ ± σ µ µ ± σ µ ± σ µ ± σ µ ± σ

Lime 0.4209 ± 0.1835 59.55% 0.3617 ± 0.1152 3.0318 ± 0.0703 0.6234 ± 9.7008 0.5250 ± 0.1041

Shap 0.3959 ± 0.2928 64.45% 0.0245 ± 0.0862 2.4677 ± 0.2074 0.6527 ± 3.8334 0.4743 ± 0.1418

IG 0.1761 ± 0.3815 73.70% 0.1774 ± 0.2505 2.1745 ± 0.4134 0.5939 ± 0.6840 0.3410 ± 0.1545

Deep Lift 0.7559 ± 0.2681 78.35% 0.0008 ± 0.0004 2.2635 ± 0.3299 0.7812 ± 25.2805 0.6754 ± 0.0897

(a) High Faithfulness (b) Max Sensitivity (c) Low Complexity (d) Relevancy Rank Accuracy

Figure 5: Quality of Explainable AI evaluation metrics distribution.

attributions.

Our analysis comparing the features identiﬁed by

the TalTech SOC analyst closely aligned with those

derived by explainers used in our LSTM model to

classify ”important” NIDS alerts. The 5 features rec-

ognized by SOC experts in Table 1 proved signiﬁ-

cant across explainers, although their order of feature

importance varied. For instance, ’SignatureIDSimi-

larity’ and ’SignatureID’, highlighted by SOC ana-

lysts, impacted the SHAP explainer for NIDS alerts.

The presence of ”SCAS” was notable in LIME, IG,

and DeepLift, conﬁrming its signiﬁcance. The im-

portance of ’SignatureMatchesPerDay’ varied among

explainers within LIME. Notably upon reviewing the

10 features highlighted by each explainer, we noticed

an overlap with the features identiﬁed by SOC an-

alysts particularly emphasizing ’SignatureID’, ’Sig-

natureIDSimilarity’, ’SCAS’ and ’SignatureMatches-

PerDay’. We assessed the quality explanation of XAI

methods, for LSTM model based alerts using metrics

based on four criteria: faithfulness, robustness, com-

plexity and reliability.

We evaluated the quality of explanations ob-

tained by XAI methods for Long Short-Term Mem-

ory (LSTM) network-based NIDS alert classiﬁcation

across 2000 data points using metrics based on four

criteria: Faithfulness, robustness, complexity, and re-

liability. Table 5 shows the results of the quality of ex-

planation for XAI methods. LSTM model prediction

probabilities were computed using the Softmax acti-

vation function. To evaluate the Faithfulness of expla-

nations, we employed high faithfulness correlations

and monotonicity. High Faithfulness of XAI methods

was evaluated by studying the correlation between at-

tribute importance assigned by the XAI method and

their impact on the model’s probabilities. A high

faithfulness correlation value suggests that the expla-

nations effectively capture the model’s behaviour and

can be regarded as faithful. Table. 2 shows the evalua-

tion results of xai methods. Mean (µ) and standard de-

viation (σ) values were calculated for the test data of

XAI computed metrics for 2000 test data points. Deep

Lift achieved the highest Faithfulness mean and stan-

dard deviation correlation values of 0.7559 ± 0.2681

for test data points. We also analyzed the monotonic-

ity of the explanation to understand how individual

features affect model probability by adding each at-

tribute to enhance its importance and observing its in-

ﬂuence on the model’s probability. By assessing the

monotonicity of the explainer, we can measure how

the explanations change monotonically with respect

to the input features. Deep LIFT achieved high mono-

tonicity with 78% (µ).

To measure complexity, we calculate the entropy

of feature attribution in the explanations. Com-

plexity measures the conciseness of explanations de-

rived by the explainer. Among xai methods assessed

by low complexity metric, Integrated Gradients (IG)

achieved lower complexity ( 2.174 ± 0.413) closely

followed by DeepLift ( 2.264 ± 0.330.)

The sensitivity metric assesses the consistency of

the explainers’ output, ensuring that similar inputs in

the feature space of model outputs have similar ex-

planations when sensitivity is low. For this metric, we

used the Euclidean distance with a radius value of 0.1

to ﬁnd the nearest neighbour points related to the pre-

diction label of an explanation which helps to identify

data points in the feature space with similar expla-

Evaluating Explainable AI for Deep Learning-Based Network Intrusion Detection System Alert Classiﬁcation

Table 3: Statistical Comparison of Explainers Across Multiple Metrics (p-values).

Metric Explainer Shap IG Deep Lift

Faithfulness

LIME L (3.34e-41) L (1.03e-134) D (5.61e-185)

SHAP - S (6.22e-91) D (1.03e-169)

IG - D (1.30e-230)

Max Sensitivity

LIME S (0.00e+00) I (1.64e-221) D (0.00e+00)

SHAP - S (1.38e-185) D (3.54e-126)

IG - D (3.29e-126)

Low Complexity

LIME S (0.00e+00) I (0.00e+00) D (0.00e+00)

SHAP - I (5.45e-146) D (1.26e-88)

IG - I (1.07e-42)

RMA

LIME S (5.12e-25) L (1.97e-80) D (6.22e-83)

SHAP - S (4.67e-155) D (6.47e-91)

IG - D (2.82e-54)

RRA

LIME L (7.61e-39) L (3.07e-210) D (0.00e+00)

SHAP - S (5.52e-155) D (3.97e-253)

IG - D (0.00e+00)

D (Deep Lift), L (LIME), S (SHAP), and I (Integrated Gradients)

indicate the better performing explainer in each pairwise comparison.

p > 0.05 — No signiﬁcant evidence against H

; H

is not rejected

0.01 < p ≤ 0.05 — Signiﬁcant evidence against H

; H

is accepted at 95% conﬁdence level

0.001 < p ≤ 0.01 — Strong evidence against H

; H

is accepted at 99% conﬁdence level

p ≤ 0.001 — Very strong evidence against H

; H

is accepted at 99.9% conﬁdence level.

nations for the predicted label. Deep LIFT achieved

Lower sensitivity with max sensitivity metric (0.0008

± 0.0004).

Two metrics, Relevance Mass Accuracy and Rel-

evance Rank Accuracy, were used to evaluate the reli-

ability of explanations. These metrics validated the

explanations by comparing them to a ground truth

mask based on features identiﬁed through collabora-

tion with an SoC analyst. For both Relevance Mass

Accuracy (0.781 ± 25.281) and Relevancy Rank Ac-

curacy (0.6754 ± 0.089) metrics, Deep lift explana-

tions were reliable. Figure. 5 illustrates the distribu-

tion of XAI metric results for 2000 data points, high-

lighting that DeepLIFT’s explanations demonstrate

high faithfulness, lower sensitivity, lower complex-

ity, and more relevance rank accuracy. Faithfulness

correlation values for DeepLIFT indicate a strong

skew towards higher levels, showing a high degree

of consistency through monotonicity. Moreover, the

entropy values of feature importance scores for IG

and DeepLIFT are more evenly spread towards the

lower end than other explainers. The sensitivity val-

ues for the DeepLIFT explainer are also more evenly

spread to lower values in maximum sensitivity met-

rics. Additionally, using Relevance Rank Accuracy,

DeepLIFT consistently achieves a high relevance rank

accuracy with less variation, centred around 0.8.

Following established practices in the statistical

analysis of XAI methods evaluation (Jesus et al.,

2021), we employed the Wilcoxon signed-ranks

test (Woolson, 2005) to evaluate the statistical signif-

icance of differences (Dem

sar, 2006) in XAI metric

scores between pairs of explainers (i.e., explainer

explainer

) for NIDS alert classiﬁcation. The null hy-

pothesis (H

) is that the explainable AI metric scores

of the explainers are equivalent, i.e., there is no signif-

icant difference between the explainers (XAI Metric

Score(explainer

) = XAI Metric Score(explainer

)).

The alternative hypothesis (H

) is that they are not

equivalent (XAI Metric Score(explainer

) ̸= XAI

Metric Score(explainer

)), indicating a signiﬁcant

difference in their explainer metric scores. XAI met-

rics used in this study are High Faithfulness, Max

Sensitivity, Low Complexity, Relevance Mass Accu-

racy, and Relevancy Rank Accuracy. This test was

conducted separately for each metric to assess the per-

formance differences among the explainers compre-

hensively.

The statistical analysis in Table 3 shows signiﬁ-

cant differences among the explainers for all metrics,

with p-values consistently below 0.05, demonstrating

strong evidence against the null hypothesis. DeepLift

explainer is better regarding faithfulness, max sen-

sitivity, RMA, and RRA when compared pairwise

(p < 0.001 for all comparisons) with other explain-

ers. The relative performance of SHAP, LIME, and

IG varies across metrics can be seen Table 3.

We have also provided a global explanation us-

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy

Figure 6: SHAP global explanation for LSTM model.

ing SHAP values for all the testing data of the LSTM

model. A higher value positively impacts the pre-

diction, while a lower value contributes negatively.

Figure. 6 shows the global explanation of the LSTM

model. The graph illustrates the average impact of

each feature on the model’s output magnitude for

the class labels, ”irrelevant” and ”important” classi-

ﬁcations. SignatureIDSimilarity, SignatureMatches-

PerDay, ProtoSimilarity and SCAS are most impact

ful features for important nids alerts. Notably, these

top features align with those identiﬁed by human ex-

pert SOC analysts. Lower-ranked features such as

HTTP-related similarities (e.g., HttpHostnameSimi-

larity, HttpUrlSimilarity) and IP-related features (e.g.,

ExtIPSimilarity) have comparatively less impact on

the model’s decisions.

5 CONCLUSIONS AND FUTURE

WORK

This research presents explainable artiﬁcial intelli-

gence (XAI) based Network Intrusion Detection Sys-

tems (NIDS) alert classiﬁcation utilizing a Long

Short-Term Memory (LSTM) model. We have show-

cased how enhancing the explainability and trust-

worthiness of AI-powered cybersecurity systems can

be achieved by clarifying the output predictions of

these LSTM models through four XAI techniques:

LIME, SHAP, Integrated Gradients, and DeepLIFT.

Our thorough assessment of the XAI framework, con-

sidering the aspects of faithfulness, complexity, ro-

bustness, and reliability, has evaluated how well these

XAI methods explain NIDS alerts. The superior per-

formance of DeepLIFT across these evaluation met-

rics underscores its potential as a preferred method

for interpreting NIDS alert classiﬁcations. Notably,

the substantial alignment between explanations gen-

erated by XAI techniques and features identiﬁed by

SOC analysts validates their effectiveness in captur-

ing domain expertise. This research makes a contri-

bution by bridging the gap between the high accuracy

of opaque machine learning models and the necessity

for transparent decision-making in cybersecurity op-

erations. By proposing a framework to explain black

box model decisions and assess XAI in NIDS appli-

cations, we provided comprehensive benchmarking

results, including evaluation metrics for developing

transparent and interpretable AI systems in crucial se-

curity domains.

REFERENCES

Alam, S. and Altiparmak, Z. (2024). Xai-cf–examining

the role of explainable artiﬁcial intelligence in cyber

forensics. arXiv preprint arXiv:2402.02452.

Arras, L., Osman, A., and Samek, W. (2022). Clevr-xai:

A benchmark dataset for the ground truth evaluation

of neural network explanations. Information Fusion,

81:14–40.

Ban, T., Samuel, N., Takahashi, T., and Inoue, D. (2021).

Combat security alert fatigue with ai-assisted tech-

niques. In Proceedings of the 14th Cyber Security Ex-

perimentation and Test Workshop, pages 9–16.

Ban, T., Takahashi, T., Ndichu, S., and Inoue, D. (2023).

Breaking alert fatigue: Ai-assisted siem framework

for effective incident response. Applied Sciences,

13(11):6610.

Bhatt, U., Weller, A., and Moura, J. M. (2020). Evalu-

ating and aggregating feature-based model explana-

tions. arXiv preprint arXiv:2005.00631.

Coroama, L. and Groza, A. (2022). Evaluation metrics

in explainable artiﬁcial intelligence (xai). In Inter-

national conference on advanced research in tech-

nologies, information, innovation and sustainability,

pages 401–413. Springer.

Day, D. and Burns, B. (2011). A performance analysis of

snort and suricata network intrusion detection and pre-

vention engines. In Fifth international conference on

digital society, Gosier, Guadeloupe, pages 187–192.

Dem

sar, J. (2006). Statistical comparisons of classiﬁers

over multiple data sets. The Journal of Machine learn-

ing research, 7:1–30.

Feng, C., Wu, S., and Liu, N. (2017). A user-centric ma-

chine learning framework for cyber security opera-

tions center. In 2017 IEEE International Conference

on Intelligence and Security Informatics (ISI), pages

173–175. IEEE.

Goodman, B. and Flaxman, S. (2017). European union reg-

ulations on algorithmic decision-making and a “right

to explanation”. AI magazine, 38(3):50–57.

Jesus, S., Bel

em, C., Balayan, V., Bento, J., Saleiro,

P., Bizarro, P., and Gama, J. (2021). How can i

choose an explainer? an application-grounded eval-

uation of post-hoc explanations. In Proceedings of the

Evaluating Explainable AI for Deep Learning-Based Network Intrusion Detection System Alert Classiﬁcation

2021 ACM conference on fairness, accountability, and

transparency, pages 805–815.

Jyothsna, V., Prasad, R., and Prasad, K. M. (2011). A review

of anomaly based intrusion detection systems. Inter-

national Journal of Computer Applications, 28(7):26–

35.

Kalakoti, R., Bahsi, H., and N

omm, S. (2024a). Improving

iot security with explainable ai: Quantitative evalua-

tion of explainability for iot botnet detection. IEEE

Internet of Things Journal.

Kalakoti, R., Bahsi, H., and N

omm, S. (2024b). Explainable

federated learning for botnet detection in iot networks.

In 2024 IEEE International Conference on Cyber Se-

curity and Resilience (CSR), pages 01–08.

Kalakoti, R., N

omm, S., and Bahsi, H. (2022). In-depth

feature selection for the statistical machine learning-

based botnet detection in iot networks. IEEE Access,

10:94518–94535.

Kalakoti, R., N

omm, S., and Bahsi, H. (2023). Improving

transparency and explainability of deep learning based

iot botnet detection using explainable artiﬁcial intel-

ligence (xai). In 2023 International Conference on

Machine Learning and Applications (ICMLA), pages

595–601. IEEE.

Kalakoti, R., N

omm, S., and Bahsi, H. (2024c). Enhancing

iot botnet attack detection in socs with an explainable

active learning framework. In 2024 IEEE World AI

IoT Congress (AIIoT), pages 265–272. IEEE.

Kidmose, E., Stevanovic, M., Brandbyge, S., and Peder-

sen, J. M. (2020). Featureless discovery of correlated

and false intrusion alerts. IEEE Access, 8:108748–

108765.

Kumar, A. and Thing, V. L. (2024). Evaluating the explain-

ability of state-of-the-art machine learning-based iot

network intrusion detection systems. arXiv preprint

arXiv:2408.14040.

Lundberg, S. M. and Lee, S.-I. (2017). A uniﬁed approach

to interpreting model predictions. Advances in neural

information processing systems, 30.

Luss, R., Chen, P.-Y., Dhurandhar, A., Sattigeri, P., Shan-

mugam, K., and Tu, C.-C. (2019). Generating con-

trastive explanations with monotonic attribute func-

tions. arXiv preprint arXiv:1905.12698, 3.

Mane, S. and Rao, D. (2021). Explaining network intru-

sion detection system using explainable ai framework.

arXiv preprint arXiv:2103.07110.

Moustafa, N., Koroniotis, N., Keshk, M., Zomaya, A. Y.,

and Tari, Z. (2023). Explainable intrusion detection

for cyber defences in the internet of things: Opportu-

nities and solutions. IEEE Communications Surveys

& Tutorials, 25(3):1775–1807.

Rawal, A., McCoy, J., Rawat, D. B., Sadler, B. M., and

Amant, R. S. (2021). Recent advances in trustworthy

explainable artiﬁcial intelligence: Status, challenges,

and perspectives. IEEE Transactions on Artiﬁcial In-

telligence, 3(6):852–866.

Ribeiro, M. T., Singh, S., and Guestrin, C. (2016). ” why

should i trust you?” explaining the predictions of any

classiﬁer. In Proceedings of the 22nd ACM SIGKDD

international conference on knowledge discovery and

data mining, pages 1135–1144.

Senevirathna, T., Siniarski, B., Liyanage, M., and Wang,

S. (2024). Deceiving post-hoc explainable ai (xai)

methods in network intrusion detection. In 2024 IEEE

21st Consumer Communications & Networking Con-

ference (CCNC), pages 107–112. IEEE.

Shin, I., Choi, Y., Kwon, T., Lee, H., and Song, J. (2019).

Platform design and implementation for ﬂexible data

processing and building ml models of ids alerts. In

2019 14th Asia Joint Conference on Information Se-

curity (AsiaJCIS), pages 64–71. IEEE.

Shrikumar, A., Greenside, P., and Kundaje, A. (2017).

Learning important features through propagating ac-

tivation differences. In International conference on

machine learning, pages 3145–3153. PMLR.

Sundararajan, M., Taly, A., and Yan, Q. (2017). Axiomatic

attribution for deep networks. In International confer-

ence on machine learning, pages 3319–3328. PMLR.

Szczepa

nski, M., Chora

s, M., Pawlicki, M., and Kozik, R.

(2020). Achieving explainability of intrusion detec-

tion system by hybrid oracle-explainer approach. In

2020 International Joint Conference on neural net-

works (IJCNN), pages 1–8. IEEE.

Tavallaee, M., Bagheri, E., Lu, W., and Ghorbani, A. A.

(2009). A detailed analysis of the kdd cup 99 data

set. In 2009 IEEE symposium on computational intel-

ligence for security and defense applications, pages

1–6. Ieee.

Tsai, C.-F., Hsu, Y.-F., Lin, C.-Y., and Lin, W.-Y. (2009).

Intrusion detection by machine learning: A review. ex-

pert systems with applications, 36(10):11994–12000.

Vaarandi, R. (2021). A stream clustering algorithm for clas-

sifying network ids alerts. In 2021 IEEE International

Conference on Cyber Security and Resilience (CSR),

pages 14–19. IEEE.

Vaarandi, R. and Guerra-Manzanares, A. (2024). Stream

clustering guided supervised learning for classifying

nids alerts. Future Generation Computer Systems,

155:231–244.

Vaarandi, R. and M

ases, S. (2022). How to build a soc on

a budget. In 2022 IEEE International Conference on

Cyber Security and Resilience (CSR), pages 171–177.

IEEE.

Van Ede, T., Aghakhani, H., Spahn, N., Bortolameotti, R.,

Cova, M., Continella, A., van Steen, M., Peter, A.,

Kruegel, C., and Vigna, G. (2022). Deepcase: Semi-

supervised contextual analysis of security events. In

2022 IEEE Symposium on Security and Privacy (SP),

pages 522–539. IEEE.

Wang, T., Zhang, C., Lu, Z., Du, D., and Han, Y. (2019).

Identifying truly suspicious events and false alarms

based on alert graph. In 2019 IEEE International Con-

ference on Big Data (Big Data), pages 5929–5936.

IEEE.

Woolson, R. F. (2005). Wilcoxon signed-rank test. Ency-

clopedia of Biostatistics, 8.

Zolanvari, M., Yang, Z., Khan, K., Jain, R., and Meskin,

N. (2021). Trust xai: Model-agnostic explanations for

ai with a case study on iiot security. IEEE internet of

things journal, 10(4):2967–2978.

ICISSP 2025 - 11th International Conference on Information Systems Security and Privacy