Using Temporal and Topological Features for Intrusion Detection in Operational Networks

TL;DR

This paper presents a novel intrusion detection approach for industrial networks using temporal motif discovery and graph-based anomaly detection, addressing unique industrial security challenges.

Contribution

It introduces an integrated method combining time series motif detection with graph analysis for improved intrusion detection in industrial environments.

Findings

Effective detection of timing anomalies in industrial processes

Identification of malicious communication patterns via graph analysis

Proposed integration enhances detection accuracy

Abstract

Until two decades ago, industrial networks were deemed secure due to physical separation from public networks. An abundance of successful attacks proved that assumption wrong. Intrusion detection solutions for industrial application need to meet certain requirements that differ from home- and office-environments, such as working without feedback to the process and compatibility with legacy systems. Industrial systems are commonly used for several decades, updates are often difficult and expensive. Furthermore, most industrial protocols do not have inherent authentication or encryption mechanisms, allowing for easy lateral movement of an intruder once the perimeter is breached. In this work, an algorithm for motif discovery in time series, Matrix Profiles, is used to detect outliers in the timing behaviour of an industrial process. This process was monitored in an experimental environment, containing ground truth labels after attacks were performed. Furthermore, the graph representations of a different industrial data set that has been emulated are used to detect malicious activities. These activities can be derived from anomalous communication patterns, represented as edges in the graph. Finally, an integration concept for both methods is proposed.

Tables

Table 1. Table 1. Research Topics Covered by the Individual Works

Subject Covered	Scientific Work
Reviews	(Garcia-Teodoro et al., 2008; Jyothsna and Rama Prasad, 2011)
Graph-based methods	(Akoglu et al., 2014; Staniford-Chen et al., 1996; Swiler and Phillips, 1998; Noble and Cook, 2003; Eberle and Holder, 2007; Pasqualetti et al., 2013; Eswaran and Faloutsos, 2018)
Graph-based and time-sensitive methods mm	(Tao et al., 2018; Akoglu and Faloutsos, 2010)
Machine learning-based	(Debar et al., 1992; Ma and Perkins, 2003; Ferdousi and Maeda, 2006)
Statistical processes	(Moayedi and Masnadi-Shirazi, 2008; Yaacob et al., 2010; Tabatabaie Nezhad et al., 2016; Yu et al., 2016)
Wavelet analysis	(Munz and Carle, 2007; Hamdi and Boudriga, 2009; Lu and Ghorbani, 2009)
Industrial Intrusion Detection	(Tsang and Kwong, 2005; Hadeli et al., 2009; Fovino et al., 2010; Regis Barbosa and Pras, 2010; Morris et al., 2012; Gao and Morris, 2014; Khalili and Sami, 2015; Caselli et al., 2015; Ponomarev and Atkison, 2016; Ghaeini and Tippenhauer, 2016)

Table 2. Table 2. Sources of the Attacks

Elem.	Sub-Process	Description	Total	No Change
P-102	P1	Pump (backup) mi	3	mmmmmmm 0
P-101	P1	Pump	2	0
MV-101 mi	P1 mmmmmmm	Motor valve	2	0
P-203	P2	Dosing pump	2	0
P-302	P3	UF feed pump	2	0

Table 3. Table 3. Detectable Points of Attacks

Elem.	Sub-Process	Description	Total	No Change
LIT-101	P1	Raw water tank level mi	7	mmmmmmm 3
P-101	P1	Pump	2	0
LIT-301 mi	P3 mmmmmmm	UF feed tank level	5	3
MV-303	P3	Motorised valve	2	0
LIT-401	P4	RO feed tank level	3	1

Table 4. Table 4. Attacks During Evaluation

Attack number	mi Source	Detectable
1	mi AIT-504	AIT-504
2	mi AIT-504	AIT-504
3	mi MV-101, LIT-101	LIT-101
4	mi UV-401, AIT-502, P-501	UV-401, AIT-502, P-501
5	mi P-602, DIT-301, MV-302 mi	DPIT-301, FIT-301
6	mi P-203, P-205	P-203, P-205
7	mi LIT-401, P-401	LIT-401
8	mi P-101, LIT-101	LIT-301, LIT-101
9	mi P-302, LIT-401	LIT-401
10	mi P-302	LIT-401, P-302

Full text

This is a preprint accepted at the 14th International Conference on Availability, Reliability and Security (ARES 2019). Please cite as indicated in the ACM Reference Format below.

Machine Learning, Graph, IT-Security, Industrial Process, Time Series

††journalyear: 2019††copyright: acmlicensed††conference: Proceedings of the 14th International Conference on Availability, Reliability and Security (ARES 2019); August 26–29, 2019; Canterbury, United Kingdom††booktitle: Proceedings of the 14th International Conference on Availability, Reliability and Security (ARES 2019) (ARES ’19), August 26–29, 2019, Canterbury, United Kingdom††price: 15.00††doi: 10.1145/3339252.3341476††isbn: 978-1-4503-7164-3/19/08††ccs: Security and privacy Intrusion detection systems††ccs: Security and privacy Network security††ccs: Theory of computation Design and analysis of algorithms††ccs: Applied computing Enterprise architectures

1. Introduction

Around the world, enterprises are benefiting from the digitalisation. The Industry 4.0 is promising reduced costs and time for acts like maintenance, set up and configuration while increasing efficiency and thus revenue. Similar to the Internet of Things (IoT) opening possibilities in end user environments, the Industry 4.0 creates new business cases (Uckelmann et al., 2011; Haller et al., 2008). However, the increase in interconnectivity and use of embedded intelligence on which the digitalisation of industry strives creates novel attack surfaces as well. In the last decades, attacks on industry have increased drastically (Duque Anton et al., 2017a). Opening network protocols that are not built for security to insecure public networks poses a threat to operation. On the one hand, criminals are targeting the cyber space, and particularly the industrial sector. Since 2007, cyber crime creates a larger revenue than drug trafficking (Dethlefs, 2015; Symantec, 2009). On the other hand, state-sponsored actors have allegedly taken a great interest in espionage and sabotage of critical infrastructure and industries, with the Ukrainian power blackouts in December of 2015 (Greenberg, 2017) or Stuxnet (Langner, 2013) being only the widely-known examples.

In order to protect industrial applications, reliable intrusion detection methods and tools are necessary. Researchers and vendors alike have taken an interest in industrial intrusion detection and prevention systems. In contrast to home- and office applications, where intrusion detection and prevention is well-established, industrial applications pose different requirements on such solutions. Legacy systems and protocols have to be integrated, as well as often application specific entities in a network. Downtimes are unacceptable as availability is the most important objective. Updates are difficult due to the distributed nature of devices. These requirements have to be considered by potential solutions. In this work, inherent features of industrial networks are made use of. First, behaviour during operation is predictable and regular. Second, the structure of a network is expected to remain stable for the most part. Those characteristics are considered in a graph-based time series approach to detect attacks in a realistic data set created by an industrial use case. First, a naive approach to time series is used to detect outliers. After that, the traffic is represented in a graph-based way as to determine cause and effect of an attack. In this application, entity-specific time series-based analysis is performed.

The contribution of this paper consists of three parts:

• •

The application of time series-based anomaly detection for detecting attacks in industrial process data is evaluated,

• •

the application of graph-based anomaly detection of Internet Procotol (IP)-based indicators of intrusions in Operation Technology (OT) networks is evaluated and

• •

the concept of a combination of both approaches is discussed

The remainder of this work is structured as follows: In Section 2, related work is presented. The data set and the industrial application it is derived from is discussed in Section 3. After that, the applied time series algorithm is introduced and evaluated in Section 4. A structural analysis of an industrial attack scenario is discussed in Section 5. Both approaches are then combined in a concept for a holistic intrusion detection approach in Section 6. Finally, the findings are discussed in Section 7.

2. Related Work

In this section, an overview of graph-based and time series-based anomaly detection in the context of intrusion detection is provided, with a focus on industrial networks. A summary of the related work is listed in Table 1.

There are reviews addressing the challenge of anomaly detection for intrusion detection. García-Teodoro et al. address the challenges of this field of work while presenting techniques and systems (Garcia-Teodoro et al., 2008). Jyothsna and Prasad provide a review of anomaly-based Intrusion Detection Systems (Jyothsna and Rama Prasad, 2011). Graph-based methods for detecting anomalies contain different approaches. Akoglu et al. provide an exhaustive overview of the different kinds of methods (Akoglu et al., 2014). They distinguish between static and dynamic graphs. In static graphs, the goal is to identify nodes or edges that are significantly different from the rest of nodes or edges respectively. In dynamic graphs, the goal is to compare the object under observation in a graph representation in different time steps. If at any time the characteristics differ, an outlier shall be detected. In 1996, Staniford-Chen et al. present an IDS for large networks, based on a graph representation of said network (Staniford-Chen et al., 1996). The network information is aggregated into activity graphs, making it feasible to detect coordinated attacks. This system is presented in more detail in a succeeding work (Cheung et al., 1996). Swiler and Phillips present a graph-based system for network vulnerability analysis (Swiler and Phillips, 1998). The general idea lies in the assignment of properties to each node and edge in the network under observation. This is mapped to possible attack vectors in the network, resulting in a probability for each type of attack on each asset in the network. Noble and Cook introduce novel methods for anomaly detection in graphs (Noble and Cook, 2003). Additionally, they present a metric for the regularity of a graph, indicating the probability of an outlier. This method is extended by Eberle and Holder (Eberle and Holder, 2007). They add types of anomalies to the model and provide methods for detecting different kinds of anomalies at different places in the graph. Pasqualetti et al. present a method specifically for detecting attacks on Cyber-Physical Systems while considering graph-properties (Pasqualetti et al., 2013). A different kind of approach for detecting anomalies in graph is presented by Eswaran and Faloutsos (Eswaran and Faloutsos, 2018). They consider dynamic graphs and look at the edges in any given time step, called an edge stream. This stream of edges is then used to detect anomalies by identifying bogus edges. Tao et al consider graph representations of online accounts (Tao et al., 2018). They extract patterns, use them to train deep neural networks such as Long Short-Term Memorys and employ these to detect fraudulent takeover of accounts.

Time series-based anomaly detection is used to detect outliers in a data series that is sequential in time. These outliers are usually anomalous in comparison to the previous values of the time series. After creating time series from a data set, there are different ways to analyse this data. Akoglu and Faloutsos combine graph-based and time series-based anomaly detection (Akoglu and Faloutsos, 2010) They consider mobile communication networks and analyse the behaviour of communication over time. If entities in the network change their behaviour, they are detected as responsible for anomalies. One way of analysing time series is by using it to train neural networks. Debar et al. present a neural network-based approach in 1992 (Debar et al., 1992). Even though they do not explicitly mention the concept of time series, they use a neural network to learn normal behaviour in a network and flag deviations as attacks. Ma and Perkins employ one-class Support Vector Machines to find anomalies in time series (Ma and Perkins, 2003). Ferdousi and Maeda employ unsupervised outlier detection techniques on time series data, namely peer group analysis (Ferdousi and Maeda, 2006).

Other than that, statistical processes are used to model the timing behaviour. Consequently, an anomaly is detected if the observed behaviour does not match the modeled values. A common approach is using Auto-Regressive Integrated Moving Average (ARIMA) to model time series behaviour (Moayedi and Masnadi-Shirazi, 2008; Yaacob et al., 2010), Tabatabaie et al. include a chaotic behaviour prediction into their ARIMA model (Tabatabaie Nezhad et al., 2016). Yu et al. present an anomaly detection scheme based on ARIMA for Wireless Sensor Networkss (Yu et al., 2016).

Apart from ARIMA models, wavelet analysis has been employed for detecting anomalies in network traffic (Lu and Ghorbani, 2009), in flows (Munz and Carle, 2007)., i.e. aggregated network traffic information and to detect Denial of Service (DoS) attacks (Hamdi and Boudriga, 2009).

Intrusion detection in the industrial domain is specific with respect to certain parameters. Legacy systems without inherent security mechanisms have to be addressed (Fovino et al., 2010; Morris et al., 2012; Gao and Morris, 2014), critical states that can have severe effects on the physical world need to be prevented (Khalili and Sami, 2015) and deterministic behaviour of processes can be leveraged to detect anomalies (Hadeli et al., 2009). Sequences are relatively unifom in industrial applications, this characteristic can be incorporated into an IDS (Caselli et al., 2015). Tsang and Kwong present an industrial IDS based on the ant colony clustering approach (Tsang and Kwong, 2005). Regis Barbosa and Pras present a novel flow-based intrusion and anomaly detection method (Regis Barbosa and Pras, 2010). Air gapped Industrial Control Systems and attacks on such systems are evaluated by Ponomarev and Atkison (Ponomarev and Atkison, 2016). Ghaeini and Tippenhauer present a hierarchical model for industrial intrusion detection to combine information from the physical, as well as the Programmable Logic Controller (PLC) layer (Ghaeini and Tippenhauer, 2016).

3. Data Set

In this work, a data set containing network and process data of a research facility is investigated. The process has been monitored for a total of eleven days, where seven days were ran as normal operation, while the four last days contained attacks. The data set used in this work is provided by iTrust, Centre for Research in Cyber Security, Singapore University of Technology and Design and is titled Secure Water Treatment (SWaT) (Goh et al., 2016; iTrust Centre for Research in Cyber Security, 2018). It consists of pcap-files containing the packets of the OT network traffic as well as csv-lists containing the sensor and actuator values at each time point. It has been widely used in scientific reserach, e.g. by Schneider and Böttinger (Schneider and Böttinger, 2018). The process contains six different sub-processes, controlled by one PLC each. In the course of the process, raw water is stored, assessed for its quality and treated by different methods. The sub-processes are:

• •

P1: Raw water storage

• •

P2: Pre-treatment

• •

P3: Membrane Ultra Filtration (UF)

• •

P4: Dechlorination by Ultraviolet (UV) lamps

• •

P5: Reverse Osmosis (RO)

• •

P6: Disposal

These sub-processes are connected as described in Figure 1.

First, the raw water is stored in a tank. It is treated by initial measures. After that, it is filtered and treated with UV light and RO. If it is sufficiently clean, it is stored in the final tank. If not, the UF and UV treatment are repeated.

Each sub-process is controlled by one PLC. These PLCs control sensors and actuators in a ring network. They are in turn controlled and monitored by Human Machine Interfaces, a Supervisory Control And Data Acquisition (SCADA) workstation as well as a data historian in a star network. A schematic representation of the communication relations can be found in Figure 2.

After seven days of normal operation, a total of 41 attacks was introduced to the process. All data is labeled, providing ground truth for the data. The operators distinguish four different kinds of attacks (Goh et al., 2016):

• •

Single Stage Single Point* (SSSP)*: Single stage attack on one point in the process, 26 instances in the data set

• •

Single Stage Multi Point* (SSMP)*: Single stage attack on multiple points in the process, 4 instances in the data set

• •

Multi Stage Single Point* (MSSP)*: Multi stage attack on one point in the process, 2 instances in the data set

• •

Multi Stage Multi Point* (MSMP)*: Multi stage attack on multiple points in the process, 4 instances in the data set

Of the total 41 attacks, 18 did not create an actual change in one of the sub-processes.

In total, 51 sensors and actuators were controlled by the six PLCs. The exhaustive list and functionality can be found in the work of Goh et al. (Goh et al., 2016). The attacks were executed on 27 of these sensor and actuators. In Table 2, the top five of affected sensors are listed, their function is described as well as the number of attacks executed on it, including the number of attacks that did not result in a change in the process.

Apart from the source, each attack could be detected at one point in the system. The top five points of detection, their descriptions, as well as the numbers of attacks that should affect them and the numbers that actually did, are shown in Table 3.

4. Time Series Analysis: Matrix Profiles

In this section, a relatively novel method for time series analysis is applied to the data set described in Section 3: The Matrix Profiles approach (Yeh et al., 2016). It was introduced by Yeh et al. in 2016 and has received many extensions since.

The concept of Matrix Profiles is to calculate the distance of any part of the time series, called motif, from any other motif in the time series. Then, the minimal distance of the given motif from any other one is used as the Matrix Profile. A low Matrix Profile indicates at least one similar motif in the time series, while a high Matrix Profile indicates an outlier. In the context of this work, outliers are of interest. A general assumption about industrial processes is a regularity in the timely behaviour. Consequently, any attack that disrupts this regular behaviour will be detected as it creates motifs that are unique. One of the non-trivial challenges of anomaly-based intrusion detection is distinguishing between non-malicious and malicious attacks. However, for the course of this work and as a generalisation, any disturbance in process behaviour is worth investigating. Only the consequence should differ, i.e. the detection of an attack should result in defense mechanisms, while non-malicious anomalies should result in maintenance efforts. Furthermore, there can be slow changes in process behaviour that are normal, as well as abrupt changes due to reconfiguration of the process. The former can be addressed by Matrix Profiles (Zhu et al., 2017), while the latter can be addressed by annotated Matrix Profiles (Dau and Keogh, 2017), as reconfigurations are expected to be known beforehand.

Matrix Profiles require one hyperparameter, $m$, the length of the motifs. This hyperparameter, however, is robust to changes. We found that the highest occurring frequency in a time series signal is a good guess. In the context of this work, two thirds of a day of normal behaviour was used as a baseline for the Matrix Profiles. Any possible deviations and irregularities in the normal behaviour are expected to occur during that time. Furthermore, about three days of process during which attacks occurred were analysed, namely the 29.12.2015 at 18:09:28 until the 31.12.2015 at 02:36:40. During this time, ten attacks, which are supposed to be detectable at twelve sensors and actuators, occur. The reason for not taking into account all of the available data lies in the size: In the course of this work, more than 230 000 time points were analysed. Each of the time points contains the values of 51 sensors and actuators. In order to evaluate the data set, sensors and actuators that are affected by the given attacks are identified and evaluated. However, Matrix Profiles do have an extension that provides multidimensionality. We used the single-dimensional approach for better visualisation and interpretation of the data, an automated approach would consider many dimensions at once.

We considered 13 sensors and actuators to evaluate for indicators of attacks:

• •

AIT-502: Sensor, measures NaOCl-level in RO-subprocess

• •

DPIT-301: Sensor, measures differential pressure in backwash-subprocess

• •

LIT-101: Sensor, measures raw water tank level

• •

LIT-301: Sensor, measures UF water tank level

• •

LIT-401: Sensor, measures UF water tank level

• •

MV-101: Actuator, controls water flow to raw water tank

• •

MV-201: Actuator, controls water flow to UF water tank

• •

P-101: Actuator, pumps water from raw water tank to second subprocess

• •

P-203: Actuator, HCl dosing pump

• •

P-205: Actuator, NaOCl dosing pump

• •

P-302: Actuator, pumps water from UF subprocess to RO suprocess

• •

P-501: Actuator, pumps dechlorinated water to RO

• •

UV-401: Actuator, removes chlorine from water

However, only AIT-502 to LIT-401 provided sensible results. The other sensors and actuators are binary in their value range. Unfortunately, Matrix Profiles do not work well with digital values. The z-normalisation is not meant for motifs with zero standard deviation which occur frequently in case of 0 and 1. Extensions are possible for boolean values encoded as binary digital, creating motifs and calculating distances based on different distance metrics. However, they are not in the scope of this work and are left as future extensions.

The attacks occurring during the evaluated time period are listed in Table 4, including the source and result of the attack.

It should be noted that attacks number 1, 2 and 3, attacks number 6 and 7, and attacks number 9 and 10 respectively are indistinguishable in the following figures due to their appearances shortly after one another. In the following figures, the attack numbers are noted next to the corresponding events. Even though the attacks are supposed to be detectable only at certain points in the system, LIT-401 can be used to successfully detect seven attacks, shown in Figure 3.

The attacks are annotated in the lower signal, one indicating an attack according to the label. This constitutes the visualisation of the ground truth for evaluation purposes. The middle line shows the Matrix Profile, i.e. the minimal distance of the motif at this position with length $m$ from any other motif. For this sensor, $m$ was set to 500 seconds. It can be seen that seven of the attacks correspond to drastic increases in the minimal distance, indicating anomalous behaviour. The sensor LIT-301 can be used to detect attack number eigth, indicated by the second right peak in the attack-line of Figure 5.

The sensor LIT-101 can be used to detect every attack except for attack number 4, indicated by the second left peak in the lower line. This is shown in Figure 5.

The sensors DPIT-301 and AIT-502 can be used to detect attacks as well, shown in Figures 6 and 7.

However, $m$ was increased in these cases. Even though it is robust to change, large differences between period and $m$ lead to suboptimal performance. The autocorrelation analysis of DPIT-301 shows a period at 2000 seconds, thus, $m$ was set to 2000 seconds. As a rule of thumb, $m$ has proven to need to at least exceed the first period of the time series data under investigation. This knowledge can be used by automated anomaly detection algorithms as well, since it is easy to compute. Starting from this value, the quality of attack detection increases. Preliminary analyses indicate that a value of $m$ of the first period or larger significantly improves the performance of Matrix Profiles. AIT-502 does not have a clear period, but the quality of detection also increases with $m$. In this evaluation, a value of 5000 seconds was chosen for $m$. AIT-504 shows constant behaviour with small deviations and one definite outlier, depicted in Figure 8. The comparably high values for the minimal distances during the training period make detection of the attacks difficult. According to the description of the data set (iTrust Centre for Research in Cyber Security, 2018), the first two attacks affect this sensor. Despite the variations, this constitutes the highest minimal distance.

5. Graph-based Analysis

The data set analysed in the previous sections contains attacks that are based on tampered process parameter. Sensor and actuator values are changed in order to disrupt the process flow. However, the attack vector that was used to gather access to said sensors and actuators is not described. Assumptions could be made, about attackers breaching air-gapped SCADA-networks of industrial applications. Since no traces of breaches can be found in the monitored network traffic, these assumptions do not aid intrusion detection mechanisms.

Due to the strict topology of industrial networks, communication patterns can be employed to detect attacks as well. Lemay and Fernandez created a set of network traffic they monitored in an emulated environment (Lemay and Fernandez, 2016). This set of network traffic contains of pcap-files containing the OT network packets for different set-ups. After simulating an industrial process consisting of circuit breakers, they introduce different kinds of attacks into the system. These attacks do not employ specifics of the used Modbus protocol, but instead are TCP/IP-based attacks conducted with metasploit. That means process-based features are of no use to detect attacks. However, network packet characteristics (Duque Anton et al., 2018b), as well as meta-data information as a time-series (Duque Anton et al., 2018a) can be used to successfully detect attacks. The number of port- and IP-pairs has a strong impact on the detection in doing so.

Three of the data sets presented by Lemay and Fernandez contain labeled malicious traffic, namely CnC_uploading_exe_modbus_6RTU_with_operate, Moving_two_files_Modbus_6RTU and Send_a_fake_command_Modbus_6RTU_with_operate. Each of them has been monitored in a network consisting of six PLCs and one Master Terminal Unit (MTU). Due to the nature of the introduced attack traffic, each data set contains communication that is not present during normal behaviour. If the communication was drawn as a graph, with the PLCs and the MTU being the nodes and any communication creating an edge between the nodes, the attacks can be distinguished as anomalous edges. This behaviour is shown in Figures 9 to 11.

In Figures 9 and 10, there are malicious packets addressed at the broadcast address. Additionally, there are malicious communication activities among the entities. This depicts the attempt to move laterally. In every scenario, PLC 1 is infected by a malware and performs different attempts to maliciously affect other entities. Furthermore, in the scenarios CnC_uploading_exe_modbus_6RTU_with_operate and Moving_two_files_Modbus_6RTU, the MTU is infected and performs malicious activities.

These kinds of attacks are relatively easy to detect, e.g. by considering the fan-in and fan-out of nodes. If it is considered as a time series that corresponds to the polling intervals, malicious changes in the behaviour are detected effectively.

6. The Hybrid Approach

As shown in the previous sections, the success of an intrusion detection approach depends on the kind of attack. If an attacker has obtained access to the control infrastructure of a process and does not try to move laterally, it is hard to detect from a network perspective. The attacker is in a position to alter parameters in a way that looks genuine. However, since the attacker tries to change the process behaviour, considering process parameters can be used to detect attacks. On the other hand, there are attacks that do not alter the process behaviour itself. Instead, network resources are used to move, extract or upload data. Even though this does not directly impact the process behaviour, this constitutes unwanted activity. It can easily be detected by considering communication graphs. For each of the data sets analysed in this work, only one of these characteristics was present and could be used for intrusion detection. The work of Lemay and Fernandez focuses on malicious network traffic that can be detected by alterations in the communication pattern (Lemay and Fernandez, 2016). However, the process remains unchanged. In contrast, the data set provided by iTrust, Centre for Research in Cyber Security, Singapore University of Technology and Design only focuses on the impact of changes in parameters on the process. Network-based lateral movement, pivoting, command and control or data exfiltration are not considered.

In combining both approaches, i.e. a time series analysis of process data and a communication pattern analysis, a holistic overview of attacks can be derived. Furthermore, the source and destination of malicious activity can be derived. The hybrid approach can be mapped on an aggregation model (Duque Anton et al., 2017b). As one of the most important requirements on industrial intrusion detection is functionality without feedback, this needs to be addressed by any IDS. Network information can be obtained from routers by mirroring the ports. The traffic should normally not exceed the throughput limit of the mirror ports. Host data, such as process information on the HMI, can be gathered there directly. The integration and aggregation of all the data can be performed in a hierarchical fashion, providing extensive information with no feedback, e.g. in employing an aggregation concept as presented by Duque Anton et al. (Duque Anton et al., 2017b). Furthermore, the concept can be adapted to an online detection fashion. This means that during operation, a data set can be created containing topology and timing behaviour along nodes and edges. Each newly introduced data point is compared to this data set, calculating an anomaly score. Since these methods do not need training as such, set up is easily done. Since the amount of data, however, is large in comparison to office Information Technology (IT) networks, storage and calculation will become tedious. Some metric needs to be employed that summarises past events in order to efficiently calculate the anomaly score.

7. Discussion

In this work, two types of intrusion detection methods were presented and evaluated. First, a time series-based analysis method, Matrix Profiles, was used to detect attacks in the process parameters of an industrial process. The process is based on an industrial environment that performs water treatment in different steps. Attacks in this work were introduced without consideration of the parameter breach, thus leaving no trace in the network communication. However, the uniform and periodic nature of industrial processes enables the analysis method to detect the effects of any attack in an unsupervised manner. Any outlier has been indicated by the application of Matrix Profiles. A major advantage is the small number of hyperparameters - one - as well as the ease of use. Furthermore, Matrix Profiles do not have a distinct training phase, making employment and set-up feasible with low effort. There are two main assumptions about outlier detection by distance calculation: First, any outlier is considered as an attack. With this method, there is no way to distinguish between misconfiguration and attacks. Second, attacks can only affect the process parameters in a unique fashion. This means that if an attack has the exact same impact on the process a second time, it will not be detected as an outlier by Matrix Profiles, as one similar motif will then create a low minimal distance. This can be mitigated by keeping track of the number of similar motifs in a time series. Furthermore, threshold calculation is a non-trivial challenge in unsupervised learning. In the evaluation, any attack could easily be detected by sight. However, formally setting a threshold after which an automated IDS triggers an alarm depends on the nature of the signal and the choice of $m$.

Second, a different data set, that has already been discussed in literature, e.g. by Duque Anton et al. (Duque Anton et al., 2018b, a), has been analysed with respect to structure. It can be seen that the attacks introduced in the data set change the topology of the communication, even though they do not impact the process behaviour itself.

Finally, the concept for integrating both methods in order to detect various kinds of attacks, based on meta-data as well as process information is presented. It should address different types of attacks and detect an attack in different stages. Usually, after a perimeter breach, lateral movement is attempted, preceded by reconnaissance. This creates anomalous traffic while not changing the process. After pivoting, an attacker could attempt to alter process parameters, changing the behaviour. If one of the stages is not detected by the IDS, detecting the attack in another stage is likely. In addition to the presented ways of detecting anomalies, attacker attribution and deception technologies, as presented by Fraunholz et al. can aid detection and identification of attackers (Fraunholz et al., 2017a, b).

Acknowledgements.

This work has been supported by the Federal Ministry of Education and Research of the Federal Republic of Germany (Foerderkennzeichen 16KIS0932, IUNO Insec). One of the data sets used in this work has been provided by iTrust, Centre for Research in Cyber Security, Singapore University of Technology and Design. The authors alone are responsible for the content of the paper.