Anomaly Detection via Federated Learning Marc Vucovicha Amogh Tarcarb Penjo Rebelob Narendra Gadeb Ruchi PorwalbAbdul Rahmana Christopher Redinoa Kevin Choia

2025-04-24 0 0 462.75KB 8 页 10玖币

侵权投诉

Anomaly Detection via Federated Learning

Marc Vucovich∗a, Amogh Tarcarb, Penjo Rebelob, Narendra Gadeb,

Ruchi PorwalbAbdul Rahmana, Christopher Redinoa, Kevin Choia,

Dhruv Nandakumara, Robert Schillera, Edward Bowena, Alex Westa,

Sanmitra Bhattacharyaa, Balaji Veeramania

aDeloitte & Touche LLP

bPersistent Systems Limited

∗Corresponding author: mvucovich@deloitte.com

Abstract—Machine learning has helped advance the ﬁeld of

anomaly detection by incorporating classiﬁers and autoencoders

to decipher between normal and anomalous behavior. Addition-

ally, federated learning has provided a way for a global model to

be trained with multiple clients’ data without requiring the client

to directly share their data. This paper proposes a novel anomaly

detector via federated learning to detect malicious network

activity on a client’s server. In our experiments, we use an

autoencoder with a classiﬁer in a federated learning framework to

determine if the network activity is benign or malicious. By using

our novel min-max scalar and sampling technique, called FedSam,

we determined federated learning allows the global model to

learn from each client’s data and, in turn, provide a means for

each client to improve their intrusion detection system’s defense

against cyber-attacks.

Index Terms—anomaly detection, federated learning, intrusion

detection systems, autoencoder

I. INTRODUCTION

As cyber-attacks continue to evolve, the intrusion detection

systems (IDS) used to detect these attacks need to stay up to

date. To stay ahead of the attackers, companies should use the

most advanced technology and work together in a way that

allows them to contribute their insights in a secure manner.

Our federated learning (FL) anomaly detection approach

provides a method for a company to detect intrusion attacks

while simultaneously contributing their insights to a global

model without sharing their individual network activity. In

a federated framework, the machine learning (ML) model is

deployed and trained on individual systems, and the weights

from the trained model are aggregated in a global model [1].

In turn, the global model will continue to adjust based on

the newly added weights and learn how to handle new data.

When paired with network anomaly detection, companies can

continuously share their insights and help the model learn to

detect more attacks.

Initial attempts at adopting FL in IDS have shown positive

improvements. However, one of the key challenges while

building an IDS is handling heterogeneous data distribution

across multiple organizations [2]. In this paper, we will discuss

some challenges that hamper interoperability of models across

organizations and how FL can be used as a bridge to overcome

these challenges. We propose training an autoencoder and

classiﬁer using our novel FL min-max algorithm and sampling

technique called FedSam.

The paper is structured as follows. First, we cover the

background information for anomaly detection, autoencoders,

and FL. Next, we discuss previous research related to our

topic. Then, we describe the data sets used in our experiments,

followed by the methods that lead to our anomaly detector via

FL. After the methods, we discuss our experimental design and

the results. We conclude with our ﬁnal thoughts and our plans

for future research.

II. BACKGROUND

A. Autoencoders and Anomaly Detection

Autoencoders are one of the most popular neural network

architectures for unsupervised anomaly detection [3]. A simple

autoencoder comprises of an encoder block (one or more layer

of neurons), a bottleneck (typically a layer with fewer neurons

than the encoder) and a decoder block (same characteristics

as the encoder), where the overall objective is to minimize

the reconstruction loss when an input is passed through this

network. Through this objective, the bottleneck layer of the au-

toencoder is able to capture the most representative features in

a lower dimensional space. When an anomalous input is passed

through an autoencoder trained on normal data, the reconstruc-

tion of the input is poor resulting in large reconstruction error.

By establishing a threshold on the reconstruction error, an

autoencoder can be used for the detection of anomalous inputs.

In an IDS, attacks are sparse and benign trafﬁc is abundant.

Autoencoders can be trained to learn diverse benign trafﬁc

and minimize the average reconstruction loss [4]. Since the

autoencoder has never encountered attack data during training,

the reconstruction loss from attack data is typically higher than

the reconstruction loss from benign data. The threshold is the

line that separates the two types of reconstruction losses such

that the amount of benign data above the line is minimized

and the amount of attack data above the line is maximized.

B. Federated Learning and Anomaly Detection

Traditionally, ML modeling techniques have relied on cen-

tralizing data from multiple sources into a single data center

to train models. However, data about different types of in-

trusion attacks are rarely located at one organization. Often,

attackers target multiple organizations, and the attack data

is spread across them. Considering the sensitive nature of

arXiv:2210.06614v1 [cs.LG] 12 Oct 2022

network related data, many organizations may ﬁnd it chal-

lenging to share data for training ML models. Consequently,

these organizations end up with an ML model that does not

achieve its maximum potential. To resolve this issue, FL

can be implemented. FL is a decentralized collaborative ML

technique [1]. Instead of aggregating data to create a single

ML model, models are trained iteratively at every node, and

the model parameters from each node are fused together using

FL fusion algorithms [1].

FL is often implemented with a central FL server node

orchestrating training rounds over multiple participating client

nodes. At the beginning of each training round, the FL

server shares a global FL model with each client node. Upon

receiving the global FL model, each client runs ML training

over the client’s local data. These clients then send the updated

model with learned parameters back to the FL server for

aggregation. The FL server collects all the updates and fuses

them by using one of the FL fusion algorithms. FedAvg is one

of the pioneering fusion algorithms [5]. Using FedAvg, the

global model update is obtained by the weighted average over

all the parameters of each client model [5]. This completes

one training round. Several training rounds are orchestrated by

the FL server until the desired performance is achieved. This

helps to ensure that client data never leaves its source location,

and it allows multiple client nodes to collaborate and build a

common ML model without directly sharing sensitive data.

III. LITERATURE REVIEW

With the constant advancements of ML techniques and

the increased availability of intrusion detection data sets,

researchers have been setting out to improve upon the current

IDS. The variety of methods used to detect anomalies with

ML have provided insights about the challenges of dealing

with cyberattack data as well as possible solutions to overcome

them.

The Canadian Institute for Cybersecurity 2017 Intrusion De-

tection System (CIC-IDS2017) and Canadian Institute for Cy-

bersecurity 2018 Intrusion Detection System (CIC-IDS2018)

data sets contain labeled network activity data for benign

and malicious behavior [6] [7]. Given the CIC-IDS data

sets contain labeled data, a classiﬁcation model is a logical

approach to determine whether the data are benign or mali-

cious. Zhou and Pezaros experimented with using 6 different

types of classiﬁcation models on the CIC-IDS2018 data set to

determine if the data are ‘evil’ or ‘benign’ [8]. The experiment

initially tested each model on individual attacks, but in the

ﬁnal experiment the team used a decision tree classiﬁer with

each of the attack types grouped together as ‘evil’ data [8].

The decision tree had an f-1 score of 1.0 detecting benign

data and 0.57 detecting the attack data [8]. The classiﬁer had

great results with detecting one type of attack, but it becomes

increasingly difﬁcult to differentiate between attacks as more

types are added.

Although classiﬁcation models have shown to be a viable

approach, autoencoders have been very successful at detecting

anomalies. Hindy et al. conducted an experiment on the CIC-

IDS2017 data set using an autoencoder with various threshold

levels [4]. The autoencoder was trained using benign data so

that the reconstruction loss would be higher when process-

ing attack data. With the optimal threshold, the autoencoder

had the following accuracies:90.01%, 98.43%, 98.47%, and

99.67% for DoS GoldenEye, DoS Hulk, Port Scanning, and

DDoS attacks [4]. These results are very promising, but

the varied accuracies based on the different threshold levels

highlights the importance of using an optimal threshold.

In another experiment, Li combines the autoencoder and

classiﬁer approaches to detect the attacks [9]. To start the

process, the normal data is sent through the autoencoder for

dimensionality reduction [9]. The data is then fed into a dense

neural network that consists of 4 layers and an output layer

for binary classiﬁcation [9]. The classiﬁer’s predictions were

then used to train and test a decision tree [9]. Along with

Li’s experiment, Rezvy et al. followed a similar approach

using an autoencoder and a classiﬁer [10]. The difference,

however, is Rezvy et al. use the autoencoder to minimize the

reconstruction error [10]. The reconstruction error is then used

as the input data for the classiﬁcation model [10]. The results

from this experiment are very promising, and the idea to use

a classiﬁer along with the autoencoder is a possible solution

to ﬁnding the optimal threshold level.

In “Chained Anomaly Detection Models for Federated

Learning: An Intrusion Detection Case Study”, Preuveneers et

al. built autoencoder based intrusion detection models using

the CIC-IDS2017 data set [11]. They partitioned data into

12 parties based on internet protocol (IP) addresses of victim

machines. The autoencoders were trained using only benign

trafﬁc from the ﬁrst day of CIC-IDS2017 simulations. In the

experiments, the authors varied the number of parties from 1

to 12. 1 represented the central training and 12 represented

the extreme case where each victim machine is a separate FL

party [11]. They observed that FL setups with more parties

required more epochs for the model to converge. In their

results, they claim that it took around 20 epochs for the central

model to converge while the 12 party FL setup took around

50 epochs [11]. While more epochs are needed, the amount of

time for each epoch reduces as each party trains local models

in parallel.

In “Federated Learning for Malware Detection in IoT De-

vices”, Marmol Campos et al. worked with malware detection

using the N-BaIoT IOT dataset [5]. They described and

compared two variations of FedAvg algorithm: Mini-Batch

Aggregation and Multi-Epoch Aggregation. In Mini-Batch

Aggregation, data at party nodes are grouped into mini-batches

for each FL round. Only a single mini-batch is used for

training, and the updated model parameters are sent back to

the FL server [5]. This process is repeated until all mini-

batches are covered. In Multi-Epoch Aggregation, the received

model is trained for multiple epochs using all the available

data at a party node before sending model updates back to

the server [5]. They described that an FL model trained with

mini batch aggregation converges better than the multi-epoch

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

AnomalyDetectionviaFederatedLearningMarcVucovicha,AmoghTarcarb,PenjoRebelob,NarendraGadeb,RuchiPorwalbAbdulRahmana,ChristopherRedinoa,KevinChoia,DhruvNandakumara,RobertSchillera,EdwardBowena,AlexWesta,SanmitraBhattacharyaa,BalajiVeeramaniaaDeloitte&ToucheLLPbPersistentSystemsLimitedCorrespondingau...

展开>> 收起<<

Anomaly Detection via Federated Learning Marc Vucovicha Amogh Tarcarb Penjo Rebelob Narendra Gadeb Ruchi PorwalbAbdul Rahmana Christopher Redinoa Kevin Choia.pdf

共8页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Anomaly Detection via Federated Learning Marc Vucovicha Amogh Tarcarb Penjo Rebelob Narendra Gadeb Ruchi PorwalbAbdul Rahmana Christopher Redinoa Kevin Choia

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: