Introduction

Visualisation has been used as a tool for monitoring networks in order to raise situational awareness levels. The static and dynamic visualisation of total and subtotal traffic information (such as bandwidth, speed and current performance) do not allow administrators to acquire a deep and clear understanding of their current network state. This is because attacks can appear like normal traffic and there are no specific rules that could enable administrators to set their network up to prevent or monitor all attacks. Furthermore, each network is unique and what is normal behaviour in one network may be anomalous in another. Therefore, network administrators need tools to provide information in a way which helps them to build a solid understanding of their network environment’s behaviour. Unfortunately, existing popular tools such as intrusion detection systems (IDS) and firewalls do not specify why and how certain events happened.

Visualisation and intrusion detection systems do not provide the protocol flow granularity required to understand how flows are behaving inside a network or why a security system generates false positive alerts or why specific alarms were raised. IDSs detect intrusions and record them to log files which network administrators must then inspect to try to understand the situation. Many IDSs send an email to the administrator for each intrusion record or incident and the volume of emails increases with the scale of the network. It is quite difficult to understand the relevance of aggregate records when receiving only the alarms for individual intrusion records. Modern attacks are sophisticated and can involve a range of of techniques and methods. Thus, real time situational awareness is required for an overall understanding of the situation especially when real time intelligence and intuitive solutions are required. The graphical user interfaces (GUIs) of today’s network monitoring and intrusion detection systems present information very superficially. For example, the time sequence of numbers of intrusions or incidents of the whole traffic domain may be visualised as polygonal charts. The operator may be required to perform many operations to explore detailed information, but in many cases network administrators are too busy to monitor the GUI. Moreover, when using visualisation tools administrators must look at a screen. Loss of concentration, visual fatigue, temporal demand and frustration increase when monitoring a screen for long periods. Extra screens will be required for additional staff. In addition, the huge volumes of data which need to be processed and presented cannot be visualised in real-time unless data reduction techniques are used.

Network measurement tools include hardware and software approaches to collect data and analyse traffic at different protocol levels. Network traffic analysers collect real-time data and perform online analysis and the majority of these systems use graphical displays to represent live traffic data.

Hildebrandt [1] proposed enhancing visualisation monitoring with sonification techniques because humans are sensitive to even small changes in the rhythms and sequences of sounds. Sonification may be defined as:

…the use of non-speech audio to convey information. More specifically, sonification is the transformation of data relations into perceived relations in an acoustic signal for the purposes of facilitating communication or interpretation [2, p. 5].

This makes sonification highly suitable for conveying information that changes over time. In the last few years there have been several attempts to develop network sonification systems in order to support network monitoring.

In order for sonification to serve network monitoring purposes administrators need to have a clear understanding of what is going on in their network environment so they can take appropriate action and prevent malicious activities and misuse of resources. The traffic volumes passing through today’s networks are huge which makes it more difficult for them to be represented visually. However, if we enable people to sense and interact with the cyber environment and let the human brain do part of the processing work and to adjust the sound generated to ease analysis this may allow more about the cyber environment to be learned.

Sonification and computer network monitoring

Sonification is introduced to enable a listener to recognise changes in activities and patterns to enhance comprehension and projection as part of the SA process. Sound allows a network administrator to continue monitoring the network while performing other tasks [7] which may, in turn, decrease frustration and visual fatigue rates. The concept of changed network behaviour as an indicator of unhealthy activity or intrusion attempts is a reasonable motive for using sonification [8]. Sonification can have advantages over visualisation in different sectors. For example, real-time sonification using parameter mapping methods is used in the health sector. A recent study showed positive results and a high potential for using real-time auditory feedback-oriented training devices for fitness training or physical rehabilitation to increase the awareness of physiological responses [9].

There is a continuing threat of intrusion, denial of service attacks, or numerous other abuses of network resources which requires the monitoring of traffic flows passing through a network [10]. The size of modern network traffic volumes makes it much harder to present real-time information visually [11]. However, there is no clear consensus yet about the pattern of cyber-attacks [12]. The assumption is that these behaviours and the rhythm associated with each type of attack should sound different or at least provide an indication of some features of any attack. Network traffic comprises packets of data which conform to specific protocols from the Internet protocol suite. The traffic protocols discussed in this paper are summarised in Table 1. For a fuller overview see [13].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Some network traffic protocols.

https://doi.org/10.1371/journal.pone.0195948.t001

Worrall [14] has described the NetSon project from its exploratory stage to a real-time sonification of network metadata. The project used the information extracted from data volumes by employing sampling techniques to extract a small group of data packets using the sFlow tool [15]. This method provides information about the network’s traffic flow rate by making a sonification of sFlow packet data of device traffic (such as printers and servers) and load balancing traffic. NetSon also provides information to identify internal and external IP addresses. This tool could be used to support network traffic measurement tools or to identify and classify IP addresses for security purposes.

Mancuso et al. [16] used sonification to help ‘cyber defenders’ to detect evidence of cyber attacks by using data collected by Wireshark [17]. The data were used offline and the source and destination IP addresses were sonified using pairs of sequential musical notes separated by 100 ms, while packet size was used to control the loudness of the sound. An experiment revealed no improvement in operator performance when using the sonifications. However, it could be argued that sonification should be tailored so that traffic with specific signatures should sound different from other normal packets, or even that sound should be generated only for malicious signatures. This might increase performance and decrease the stress of the operator.

Vickers et al. [4, 11, 18] applied sonification to the inherent self-organised criticality observed in network traffic. Standard packet capture tools were used to gather network traffic which was then passed to the socs (self-organised criticality sonification) system which sonified the log returns of packet sizes at regular user-specified intervals. The extracted log returns provide information about the behaviour changes in the network. Knowledge of this behaviour could be used to detect unwanted behaviour. This system has potential to support both network traffic measurement and intrusion detection tools.

Wolf and Fiebrink [19] developed SonNet, a programming interface for sonifying computer network data. The prime motivation behind SonNet was to lower the practical barriers for artists and sound designers interested in accessing network data to create music. SonNet involves packet sniffing and offers network state analysis and easy access to computer network data for composers. The tool supports the sonification of data using the UDP and TCP protocols. SonNet extracts network data at various levels from packet level information to network state information. Level 1 contains information about a single packet, level 2 contains information generated by computing and analysing the single packet information, and level 3 contains information about multiple packets. In a similar vein, Rutz et al. [20] introduced the SysSon platform for developing sonification applications for different types of users from domain scientists to sonification researchers, composers and sound artists.

InteNtion (Interactive Network Sonification) [21] is a project targeted at mapping network traffic activities to a musical aesthetic. Network traffic data is converted into MIDI messages and then sent to dedicated synthesisers to generate a dynamic mix of sounds as an interactive soundscape. The system uses Internet protocols including TCP and UDP segments, using very low-level packet information such as packet size, source and destination IP addresses and type of service. The work is still considered to be experimental and the system needs more development and better mapping to support network traffic monitoring. However, it provides an innovative way to monitor a network by using the entire data flow to create music.

Earlier work done by Ballora and Hall [22] explored the detection of intrusion signatures and patterns using human aural and visual recognition abilities to detect intrusions in real-time. IP addresses and return codes were used to generate sound as an informative and unobtrusive listening environment to develop web traffic SA. Ballora et al. [8] conducted another sonification experiment with a computer network based on socket connections using information such as the date and time of exchange and the sender’s and receiver’s IP addresses and port numbers. Ballora et al. [23] have also described the use of sonification in the detection of anomalous events. Sonification should enable the listener to differentiate between normal and anomalous network behaviour and to develop an understanding of what is actually happening in the network.

Kimoto and Ohno [24] introduced the Stetho network sonification system which was aimed at system administrators. NetSound was built on top of Stetho as a tool for end users. Stetho used network traffic information to generate sounds which provide the network administrator with information about the traffic. Stetho reads the tcpdump [25] commands, then uses them in regular expressions to generate corresponding MIDI events. Stetho processes each packet in the traffic. However, Stetho failed to detect all events and intrusions. Delays in sound generation and poor MIDI messages generated further problems.

Chafe and Leistikow [26] developed a tool for the measurement of round trip time when using a sequence of standard ping utility events to gather information about the quality of service of a network path, such as packet loss. They discussed the need to evaluate paths which carry interactive media streams in collaborative environments. They designed a stream-based method for the direct display of the critical qualities to the ear by continuously driving a bidirectional connection to create sound waves. They changed the network path to acoustic medium which their probe sets into vibration. Temporal levels of musical foreground, middle-ground and background are heard in the melodies generated from correspondence data.

SoNSTAR

The sonification of high-speed computer networks demands both high throughput and flexibility to handle and recognise new threats. It is possible that sonification is a viable solution to this problem and could allow an administrator to listen in real-time to the state of each traffic flow. As a solution to these problems and issues, we propose SoNSTAR—Sonification of Networks for SiTuational AwaReness—to be used by network administrators as a monitoring tool to facilitate the acquisition and maintenance of network situational awareness. SoNSTAR would assist with the maintenance of security, awareness of anomalous events such as attacks, maintenance of network health through monitoring and tuning, and increasing the understanding of the cyber environment which is vital for network management the use of diagnosis to support the recognition phase in the situational awareness process.

A traffic flow is a flow of packets in a single connection between a source computer and a destination [27]. A single flow can be identified within a certain time period by its source and destination IP addresses and ports and its protocol type and layer (such as TCP, UDP and ICMP). As part of our technical solution, we have created a new flow type called IP flow which is identified within a certain time period by its source and destination IP addresses and protocol only (no port information). Thus, SoNSTAR uses these two flow types (traffic flow and IP flow).

SoNSTAR uses events to generate sounds. A flow event is a change in the behaviour or operation of a flow (traffic or IP). A single event represents a combination of features of a flow while a set of events represents flow behaviour which, in turn, represents the state of the network traffic.

In the TCP protocol, the packet header contains nine control flags, six of which (FIN, SYN, RST, PSH, ACK and URG) are used by SoNSTAR. Values of 1 and 0 denote whether a flag is set or unset, and the packet’s type is determined by those flags that are set. A packet’s type determines its role and function within the network traffic. Therefore, SoNSTAR collects counts of each packet type for both traffic- and IP-flows. A flow’s status is determined by the respective packet type counts. SoNSTAR allows its user to listen to the status of the flows in the traffic by playing sounds that represent the flow behaviours.

Thus, SoNSTAR makes information about traffic perceptible allowing the network administrator to make decisions about network operation on the basis of recognising the sounds that describe the network environment. SoNSTAR allows users to set specific sounds for different flow status types and to tune the thresholds for triggering the sounds. What makes SoNSTAR distinctive compared to other available tools is that it allows the user to monitor general and specific behaviours in a human understandable form.

When designing a sonification system with the purpose of monitoring a system or network activity to gather administratively useful information, the design will involve a number of conditions and requirements. The sound has to support extended periods of listening, changes in status have to be easily grasped and accidental events have to be immediately noticed [24]. SoNSTAR sonifies each flow in a connection and collects information about the connection state by periodically gathering online flag information from each flow. Traffic features are extracted from the flag information aggregations and SoNSTAR then represents these features in a soundscape.

The term “soundscape” was introduced by Murray Schafer [28] and describes the sonic properties of landscapes. Sounds are a continuous and active property of all landscapes and soundscape ecology is generated from the sounds and spatial temporal patterns as they occur in a landscape environment, where each sound has special ecological characteristics [29]. In SoNSTAR the network environment is transformed into an acoustic environment as a soundscape and the combinations of sounds represent the current state of the network, just as the combinations of sounds in a landscape provide information about what is happening in the environment.

SoNSTAR transforms the network environment to the soundscape of a forest (though it is fully configurable and allows any other soundscape to be used as desired). Just as a person would be able to infer information about what is happening in a forest by the sounds they hear, sounds in the soundscape represent events and unexpected or particularly loud sounds can draw the listener’s attention to traffic behaviour that is out of the ordinary.

Using recorded sounds in sonification can be difficult, as there are limitations on how recordings can be used to represent traffic while still sounding realistic [30]. However, the use of recorded sounds is better than synthesised sounds, because it enables users to link events to familiar and understandable sounds. Sounds from a natural environment such birds tweeting or animal sounds are easier to describe than artificially synthesized tones which may rely on specific terminology such as frequency and timbre [31]. The sounds provide us with immediate awareness of the types of events that are happening:

Modern cognitive science believes that to be able to read sound in this way, listener must have some inner understanding of how the properties of physical events are reflected in the sounds they make [32, p. xviii].

Therefore, a monitoring operator requires a good understanding of communication protocols and theoretical and practical knowledge about the expected behaviour in computer networks. SoNSTAR allows the user to make a relation between the meaning of the recorded sound and the event mapped to within the network environment. For example, a forest on a normal day will produce sounds of typical birds and animals, perhaps with a very light breeze. These events can be used to describe the normal state of a computer network environment, while human-made sounds, the sounds of predators and changes in weather (such as rain and thunder) can be used to represent abnormal or malicious network activity. SoNSTAR sonifications are generated using an event mapping method based on flag state information collected from each TCP packet for each flow in the network. This specialised abstraction of network features is extracted from the raw flow packets and transformed into classified sound groups of natural and human-made sounds.

SoNSTAR and network traffic monitoring

Commonly, administrators try to look directly at network traffic to understand it using tools such as Wireshark [17]. Network traffic volumes can be huge and the majority of the traffic involves normal data packets travelling between legitimate users on the network or across the Internet. TCP/IP packets carry control flags to allow the data to be received in sequence and to protect it from loss. In TCP/IP, if receipt of any packet is not confirmed by the destination, it will be sent again. In contrast, in the UDP protocol any packet sent will be considered as received and packets will be processed in the order they arrive regardless of whether the routing has caused them to be received out of sequence. In TCP/IP approximately 30%–40% of traffic concerns packets which are very important to administrators for enabling them to understand immediately what is happening in their network environment [33]. This means that the TCP/IP control packets SYN, SYN ACK, ACK, FIN and RST provide most of the information about network traffic state. UDP packets have to be monitored in such a way that allows administrators to recognise the current state. TCP/IP traffic accounts for more than 85% of packets entering and leaving a system or computer network [33, 34]; therefore, TCP traffic is considered a priority.

Network administrators typically identify anomalies in traffic from two sources. The first is simple network management protocol (SNMP) data from queries to network nodes. However, the data collected from the SNMP management information base (MIB) is wide ranging, and contains activity statistics such as total packets transmitted at a node. This source can only provide statistics about volumes of packets and bytes which provide useful information but cannot be used to understand the behaviour in the traffic flows and connections in the network. The second source is the monitoring of end-to-end packets, flows or connections. This data contains protocol-level information. This second source is typically used by intrusion detection systems. These two sources offer a practical base for the identification and recognition of anomalies as part of situational awareness [35].

SoNSTAR uses the second source and collects data by sniffing the traffic passing through a switch or a router from the mirroring outlet in real time or by reading stored PCAP files captured by any other available packet sniffing programs. The sniffer act as a sensor that periodically collects traffic information.

Several types of monitoring systems use network usage patterns for detection, measuring usage and summarising usage statistics based on user-defined parameters, and contrasting measurement aggregates with predefined thresholds and then responding when thresholds are met or exceeded or following queries from a security analyst [36].

Most detection methods (especially IDS) depend on packet headers or the payload or a combination of both to detect attacks and malicious activity. In anomaly-based systems analysis of the packet’s payload is used to differentiate between normal traffic and anomalous activity. Signature-based systems rely on matching patterns with a database of the signatures of known attacks. The advantage of anomaly-based systems is that, unlike signature-based systems, they can detect attacks without any delay since new attacks can be detected as soon as they happen, while signature-based systems cannot detect novel attacks and can only match against known attack signatures [10, 37]. While anomaly-based systems can detect novel attacks they generate more false positive results and so risk blocking legitimate activity. Identifying the state of traffic from encrypted applications is a critical issue for numerous network tasks. In-depth packet inspection requires decryption in most cases, and this would affect any detection mechanism especially when trying to operate in real time [38–40].

SoNSTAR uses packet header information to generate sounds which periodically represent the status of aggregated packet information for multiple flows in the network. It is an anomaly-based system which generates different sounds according to the network state. This method can be used to provide a general or specific sonic representation of the traffic behaviour. Any changes in sound combinations then represent a new state or behaviour. An advantage of this approach is that an administrator using SoNSTAR can interact with the system and change and create the features to be sonified and assign sounds to those features. SoNSTAR is an additional tool that enables administrators to discover changes in and learn more about their environment in a way that enables the human mind to comprehend the mechanism of these changes and their causes.

A security system using real-time monitoring for situational awareness has to show changes in flow and connection states as they happen and provide an indication to the administrator about immediate events. SoNSTAR targets this type of monitoring to support existing security tools, acting as an additional tool aimed at raising situational awareness levels.

SoNSTAR design

Computer network defence requires analysts to detect both known and novel forms of attack in massive volumes of network data. Visualisation tools would potentially assist in the discovery of suspicious patterns of network activity, but few analysts so far are leveraging sonification technologies in their current security practice. We have designed SoNSTAR to suit the work practices and operational environments of those analysts.

The work’s novelty consists in sonifying in real-time the extracted features of network traffic based on the control flag status of the packet header and the techniques developed to handle the interaction of the user with the system with the aim of increasing situational awareness levels. SoNSTAR inspects the flag statuses of each packet in the flows and extracts features by periodically counting each packet type and counting the number of flows and then uses this information to control the resulting soundscape. This results in a system that is complementary to and more informative than visualisation methods, but which can provide only limited goal-oriented information. This type of sonification, which allows the representation of large traffic volumes by representing traffic flow and IP flow states to reduce the amount of traffic information presented to the user, has not been done before.

In many systems, changes in performance could be used to indicate the vulnerability or robustness of a computer network [41]. Equally, changes of sounds could be used to indicate changes in network behaviour. The first goal of the design of the sonification system as part of the situational awareness process is either to monitor network assets or the network gateway and to find a way to sonify network component activity and traffic behaviour to enable the listener to detect any misuse or anomalous behaviour. This anomaly detection approach must first determine the normal behaviour of the object being monitored, and then use deviations from this baseline to build experience and knowledge to detect and identify possible malicious activities.

Monitoring tools try to present administrators with a complete representation of their complex network. Better network monitoring tools should allow administrators to perceive changes in their network in order to allow them to react immediately, and learn and understand more about the cyber environment. A real-time sonification monitoring tool should be able to do or assist with the following:

• Identify and recognise malicious traffic: Malicious traffic such as probes and denial of service attacks should be indicated.
• Provide information about incidents or changes in behaviour: An incident or change in traffic behaviour should be reported to allow the user to recognise which flows are malicious.
• Represent network behaviour sonically and in a non-fatiguing and non-annoying way: Sounds representing states have to be easily recognised and linked together by the user to allow comprehension as part of the situational awareness process.
• Offer practicality: Use of the system should be convenient for both incident response and real-time monitoring.
• Indicate compromised machines: A machine compromised by a hacker or malicious software such as worms or viruses should be indicated when ever possible.
• Offer high throughput and flexibility: The system should be able to handle large amounts of data in a timely manner and its operation should not be CPU-intensive.

Design solution

This section considers a design that can fulfil the requirements of real-time monitoring for situational awareness. The practical issues associated with the selected design are also discussed.

The SoNSTAR architecture diagram is illustrated in Fig 1. The system is implemented in Python using the pcapy and dpkt libraries and Max/MSP [47]. The Python engine captures and processes the packet information and passes data to the Max/MSP patch which generates the audio.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. SoNSTAR architecture.

The major components of the system.

https://doi.org/10.1371/journal.pone.0195948.g001

SoNSTAR uses time a window period to arrange and control the timing of the operation of each process within the system (see Fig 2). SoNSTAR reads packets and unpacks them and filters the TCP packets and extracts counts during time window X. At the end of each time window, features are combined to generate higher-level aggregate features. The selected features are then represented as recorded sounds. These sounds are played during the next time window Y.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Time window processes.

SoNSTAR aggregates flow data across time windows. This figure shows the process timing and sequencing across two time windows, X and Y.

https://doi.org/10.1371/journal.pone.0195948.g002

The time window is user defined and can be any value above 5 s which is the minimum required to allow SoNSTAR to collect sufficient packets to count target events. During development SoNSTAR was applied to a variety of network traffic (both datasets and live traffic) and a 20 s window was found to be a good default value to enable enough features to be collected for setting distinctive events. Because the size of a network determines only the number of flows passing through it, it does not affect the time window duration or SoNSTAR’s ability to collect features and handle event conditions. Thus, it is left to the user to decide which time window most suits the events he or she wishes to target. Moreover, the time window value would affect the thresholds of event conditions which allow the user to develop an intuitive sense of the link between features and events. The only constraint is that a very large network might require more than one instance of SoNSTAR to be installed on lower hierarchy routers, each representing a smaller sub-network, to avoid overloading the processing capabilities of the machine on which SoNSTAR is installed.

The main SoNSTAR algorithm is shown in Algorithm 1. The SoNSTAR system comprises five blocks described below.

Algorithm 1 SoNSTAR’s main algorithm

Set Timewindow period

Sniff packet and Get start time

if Packet ⩵ arrived then

 Unpack ethernet header

 Extract EtherType

 if EtherType ⩵ 0x0800 or 0x86DD then            ⊳ IP packet

  Unpack IP header

  Extract source and destination addresses

  Extract transmission protocol

 else

  Get next packet from the sniffer

 end if

 if Protocolnumber ⩵ 6 then                 ⊳ TCP packet

  Unpack TCP header

  Extract flags information according to incoming or outgoing

  Count flags status according to incoming or outgoing

  if Timewindowperiod ⩵ finished then

   Extract current flag’s features

   Extract new features from Features Combiner

   Apply thresholds to selected features

   Send messages to Max/MSP for sonification

  end if

  Get next packet from the sniffer a new Timewindow started

 else

  Get next packet from the sniffer

 end if

else

 Get next packet from the sniffer

end if

Max/MSP Patch

if messages ⩵ arrived then

 Play sound of similar messages once

end if

Sniffer.

The main input to the system is the raw traffic packets passing (incoming and outgoing) through the network and the Sniffer reads these packets in real time.

Filter.

The Filter unpacks each ethernet frame, extracting the packet header information, and sending only TCP packets to the Feature Extractor. A TCP/IP packet has an EtherType value of 0x0800 or 0x86DD (denotes IP protocol) and a transmission protocol number of 6.

Feature extractor.

Next, the Feature Extractor picks up each TCP packet, checks the flag values, and determines the packet type. If this flow has not been seen before it creates a new Traffic flow and IP flow and sets the counter for the current packet type to 1 for each flow. If the flow already exists the feature extractor increments its packet type counts by 1 according to the packet’s direction (incoming or outgoing). This update happens for both flow types (traffic flow and IP flow). At the end of the time window, the set of traffic flows with their packet type counts and the set of IP flows with their packet type counts, in addition to number of traffic flows and number of IP flows are passed to next stage. The full feature sets for both flow and IP-flow packets can be found in the supplemental material (see S1 Appendix).

At the end of each time window SoNSTAR creates two log file reports consisting of the entire set of Traffic- and IP-flows and their respective packet type counts for any post-hoc inspection and review that may be required.

Feature combiner.

The Feature Combiner enables the user to create new features by adding or subtracting particular flags (see Table 2 for some examples). This enables the user to target specific flow events. Some of these combinations could be set according to user needs and understanding of the TCP protocol behaviours and rules. Some could be built over time while listening to and learning about the network environment’s behaviours and sounds.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Feature combinations.

https://doi.org/10.1371/journal.pone.0195948.t002

For example, TCP requires the use of specific mechanisms to establish connections between source and destination hosts. An established process is called the three-way handshake. The first step in the process is that the source (S) sends to the destination (D) a TCP packet with the SYN flag set. Next, D replies to S with a packet with the SYN and ACK flags set to acknowledge and accept the connection. Finally, S sends a packet to D with the ACK flag set indicating acknowledgment of the agreement. In this way the handshake process is successfully completed and the connection is established. After the exchange of data and at the end of the connection, either side will terminate the connection by sending a TCP packet with the FIN flag set [48]. Therefore, each flag’s status gives us information about the flow and changes in flag status represent what is happening in the network.

At this stage of SoNSTAR design we have created some new features from previous IP flow features (see S1 Appendix) provided by the Feature Extractor (see Table 2). All of these features are now available for sonification.

Sonification.

The final block in the system is Sonification. To make sense of the sonification we must assign sounds according to event conditions and thresholds and according to the understanding of flag status mechanisms for both flow types. Knowledge of these events could be learned over time while listening to the network environment, tuning the thresholds and experimenting with conditions to target particular behaviours and exploring log files.

Through development of this design recorded natural sounds have been assigned to various features to create a network soundscape environment. By operating SoNSTAR and listening to sounds and manipulating event conditions and tuning thresholds, new events and feature combinations can be defined (such as those new features listed in Table 2). Threshold values could vary according to the characteristics of the network being monitored.

Of the many features that could be monitored for intrusion detection purposes, some are truly useful and some are less significant, and may indeed be useless. A standalone IDS might generate many false positives or could ignore an anomaly (false negative) depending on its settings. There is no clear analytical model that provides the basis for a mathematical formula to precisely describe the input-output relationship [49]. Therefore, using SoNSTAR would provide that missing understanding of the decisions made by an IDS and allow its user to gain knowledge through monitoring the real behaviour and events of the flows within the traffic.

Every network is a unique environment. Relationships between features are important when applying sounds to the events chosen. This is what gives SoNSTAR a real advantage in exploring a network environment because the understanding of the traffic environment can be improved by taking into account feature relations. The idea behind using different recorded sounds from nature and human-made sounds to represent the network environment is to transform the experience into an interactive soundscape environment. The sounds generated express the behaviour of flows and their deviations from the normal state in order to increase situational awareness.

Representational techniques

Sonic representation is a challenge because of the huge volumes of traffic passing through each connection in the network. Each connection has a high potential number of flows depending on the nature of that connection and its purpose. SoNSTAR reduces the complexity of representing huge volumes of traffic by two methods. The first considers IP flows rather than traffic flows. A number of traffic flows could exist between any two hosts as each traffic flow is specific to a single port number. IP flows are not concerned with port numbers so the number of flows between any two hosts is reduced to one for sonification purposes (see Fig 3). In the second method SoNSTAR maintains counts of the packet types for each traffic flow to update the soundscape at a user-specified interval. Since network traffic consists of a number of flows which can be similar in their condition, so similar flows can be expressed once so that there is no repetition of the same sound. By doing this we have reduced the number of flow events that need to be sonified.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Conflation of multiple traffic flows to one IP flow.

Seven traffic flows between different ports on the same sending and receiving hosts are reduced to a single IP flow.

https://doi.org/10.1371/journal.pone.0195948.g003

Recorded sounds (such as birds or rain) represent discrete events by playing a single natural sound every time the event occurs. The sounds chosen are diverse in nature and easily distinguishable by the listener.

Tuning the system.

One begins to tune SoNSTAR for a particular network by starting with the three-way handshake mechanism and assigning it to a chosen sound. Then, each flow event of interest is mapped to a sound and then its frequency of occurrence is listened to over time in order to get a sense of its impact on network behaviour. The event’s feature threshold value can then be adjusted to suit. It was noted during development that certain events tend to occur normally in every network or dataset. Network mechanisms and activities which are confirmed as normal events were mapped to sounds from a forest birds collection. Forest birds were used because they represent the normal state of a forest. Sounds that do not belong to the normal state of a forest were then used to represent rarer, unusual, or anomalous events. Fig 4 shows an example sonification of IP flows to represent network traffic state. (Listen to the file S1 Audio of normal traffic sonification in the supplemental material).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. IP flow representation.

Illustration of multiple IP flows containing a range of different events and even combinations are mapped to different sounds resulting in a sonic representation of the overall traffic state.

https://doi.org/10.1371/journal.pone.0195948.g004

Events which are outside the normal range are represented according to the main flag type that caused that event. Sound representation is divided into five categories. The first category of network states represents ongoing events related to SYN or SYN-ACK packets (or combinations thereof) and is represented by weather-related sounds of rain or water. For example, the soundscape changes from rain to heavy rain to rain and thunder according to the number of packets that caused the event.

The second category represents ongoing FIN, ACK, URG, PSH or NULL packets (or combinations thereof) and is represented by animals or unusual birds. The third category represents ongoing RST events and is mapped to wind sounds. For example, when any host sends a high number of RST packets the sonification reflects the change in network state by playing a wind on grass sound; if the RST packet changes usual behaviour in relation to SYN and FIN packets, a heavy wind sound is played. The fourth category represents ongoing events related to traffic- or IP-flow counters and is represented by sounds of fire in the woods. The fifth category represents ongoing events confirmed as normal conditions and is represented by usual forest birds forming ongoing background sounds. Fig 5 shows an example of event representation in SoNSTAR.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Event representation.

Illustration of different events (identified the main flag type) being mapped to discrete sounds the SoNSTAR soundscape.

https://doi.org/10.1371/journal.pone.0195948.g005

For a better representation, incoming and outgoing events of the same type are represented such that incoming events are given more worrying and louder sounds (more dangerous- or urgent-sounding versions of the sounds) than outgoing events which are quieter which are mapped to non-alarming animal sounds. Furthermore, it was observed that several events tend to occur together or in specific sequences for particular types of attack. Therefore, their sequenced sounds were examples of behaviours that were learned as SoNSTAR was used to begin exploring network traffic. It is posited that the information about network traffic provided by SoNSTAR can assist with the recognition of anomalies, both of known and unknown (not previously encountered) types.

Sound design and representation depend very much on personal taste and targeted behaviour. SoNSTAR provides the user with a choice of sound sets (e.g., forest, weather, and animals sounds, or even human-made ones) and assigns sounds according to the event features the user wishes to monitor.

SoNSTAR features sound mappings

The features used for sonification are aggregation counts of the flag status of each flow type in the traffic. For each feature thresholds are set such that sounds are generated only when the counts exceed the threshold. Users can select the thresholds appropriate to their network environment. A set of default mappings was created based on an understanding of TCP protocol theory and running SoNSTAR several times whilst carrying out simulated attacks in order to learn about traffic features. The thresholds used do not represent a priori fixed rules. However, experimenting with these thresholds requires an understanding of the flag relations in the TCP protocol. Network traffic is not static and what can be normal traffic behaviour in one context could be malicious elsewhere, and thus the expected numbers of flows could vary depending on the purpose of the network. The default event-to-sound mappings are listed in Table 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Feature-to-sound mappings.

https://doi.org/10.1371/journal.pone.0195948.t003

SoNSTAR interactive sonification

SoNSTAR is an interactive sonification system. Users may change the time window period, manipulate features and thresholds and re-assign sounds, and then restart with the new settings online. The level of each event sound can be adjusted independently with a slider control and can even be muted if desired. Any sound can be assigned to any chosen flow event in real time enabling the user to re-design the sound environment completely. Fig 6 shows the SoNSTAR interactive sonification model.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Interactive sonification.

Model showing the interactive nature of the SoNSTAR sonification.

https://doi.org/10.1371/journal.pone.0195948.g006

Users interact with SoNSTAR according to their understanding of the sound generated by the network traffic environment so as to increase their situational awareness. SoNSTAR enables the user to interact immediately with the system and its traffic to identify anomalous behaviours. Hunt and Hermann advise that sonification designers should respect ‘the bindings between physical actions and acoustic reactions that we have been familiar with since birth’ [50, p. 295]. In a network environment this could mean that we would expect the sounds to change when the system is under attack and we expect networks to behave differently when they are under more stress. SoNSTAR uses multiple natural and man-made sounds to create the soundscape environment. When choosing the sounds, the natural reactions of users to the sounds is taken into consideration in order to allow users to sense and feel the network environment in relation to their own experience in the real world. SoNSTAR allows users to change sounds and create their own preferred acoustic environment in order to enable them to choose the most suitable sounds which convey to them the state of the network in a maximally meaningful way. SoNSTAR transforms all of the network traffic into a rich auditory field that envelops the listener in a goal-driven exploratory methodology where the network traffic is first filtered and the user is left only with the specific features that they chose.

Experimental work and results

A user study was conducted to investigate the monitoring of network behaviour by participants using SoNSTAR and, in particular, to evaluate SoNSTAR as a complement to existing system security tools. Three experimental conditions were investigated: 1) audio feedback only using SoNSTAR, 2) visual feedback only using the Snort IDS, and 3) audio and visual feedback together (SoNSTAR and Snort).

SoNSTAR’s current design is able to extract TCP, UDP and ICMP protocol packet information. In this experiment only TCP and ICMP packet header information was extracted with ICMP packet data being used to detect ping activities. The sound of a woodpecker sound was assigned to ICMP ping activity.

Network design

The experiment was conducted using two virtual networks running on the Virtualbox software. The first network was installed on a macOS 10.10.5 workstation with a 3.7 GHz quad-core processor, 16 GB 1866 MHz DDR3 ECC RAM and a 27-inch (2560 x 1440) display. This virtual network comprised four machines (Ubuntu 64-bit, Windows Server 64-bit, Kali Linux Debian 64-bit and macOS 10.11) in addition to the host machine.

The second virtual network was installed on a MacBook Pro running macOS 10.10.5 with a 2.5 GHz Intel core i7 processor, 16 GB 1600MHz DDR3 RAM and a 15.4-inch (2880 x 1800) Retina display. This network contained three machines (two Kali Linux Debian 64-bit installations and a Fedora 24 64-bit machine) in addition to the host machine. Fig 7 shows the network environment design.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Virtual network environment.

The virtual network environment design used in the experiment.

https://doi.org/10.1371/journal.pone.0195948.g007

These two virtual networks were connected through a router provided by Northumbria University. SoNSTAR and the Snort IDS were installed on both networks allowing each network to attack the other, and each machine to attack the other machines within its own local virtual network.

Experimental design

Each participant performed a network monitoring task under each of the three experimental conditions (audio only, visual only, audio-visual). Each task required participants to detect either 3 or 4 out of 7 attacks.

The participants were assigned randomly to use Snort or SoNSTAR (five participants each) and then to use them together. At the end of each task performance was calculated based on the number of true positive (TP), true negatives (TN), false positives (FP)and false negatives (FN), where:

• TP: the number of events which are correctly identified. The case was positive and was detected by the user as positive.
• FP: the number of events which are incorrectly identified. The case was negative but was detected by the user as positive.
• TN: the number of events which are correctly rejected. The case was negative and was detected by the user as negative.
• FN: the number of events which are incorrectly rejected. The case was positive but was detected by the user as negative.

These variables are then used to calculate several metrics to assess the effectiveness of SoNSTAR as follows.

The recall metric (also known as the true positive rate) indicates the proportion of positives which are correctly detected by participants and is given by:

The precision is the number of true positives amongst all the reported positives:

The F-measure is a weighted harmonic mean of the precision and recall [51, p.1147]:

The accuracy metric indicates the proportion of correct identifications of all instances:

The true negative rate (TNR) indicates the proportion of negatives that are correctly identified, such as the percentage of network events which are correctly identified as not occurred.

The false positive rate (FPR) indicates the proportion of positives that are incorrectly identified, such as the percentage of network events which are incorrectly identified as occurred.

The false negative rate (FNR) indicates the proportion of negatives that are incorrectly identified.

Snort’s detection rules were set to the defaults provided by the snort.conf file. SoNSTAR was set to the sound mappings presented in Table 3.

Four categories of behaviour were used in this experiment as follows:

• Traffic: using the Internet, such as playing a YouTube video.
• Ping: using an ICMP ping.
• Port scan: four types—SYN, Null, Xmas and FIN port scans.
• DoS, DDoS including first, SYN flood as type; and second, DDoS using spoofed IP addresses performed from the three machines in the virtual network.

These behaviours were performed using a normal terminal, Nmap scanner and Hping3 commands. The supplmental material contains the files S2, S3, S4, S5, S6, S7 and S8 Audio which are the SoNSTAR sonifications of the attacks used in this experiment.

Results

Several results are extracted from the questionnaire data as follows.

TP, TN, FP and FN results.

The results for the three conditions are shown in Table 4 as extracted from the questionnaire data. The results were calculated for the three conditions to assess SoNSTAR’s capabilities as part of the situational awareness process. Based on these results, various metrics are calculated to evaluate the SoNSTAR sound design and the usability of the system.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. TP, TN, FP and FN.

https://doi.org/10.1371/journal.pone.0195948.t004

The metrics calculated from the base variables are shown in Table 5. The results show a maximum recall of 100% for the three state conditions. Meanwhile, the TNR was higher when using SoNSTAR (89.19%) compared to Snort (81.58%). However, when participants used both together this rose to 95%. The FPR was higher when using Snort (18.42%) than SoNSTAR (10.81%). However, when participants used both together this decreased to 5%.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Evaluation results.

https://doi.org/10.1371/journal.pone.0195948.t005

Accuracy was calculated for the three state conditions used in the experiment. Accuracy of recognition was highest when using both Snort and SoNSTAR together at 97.14%. SoNSTAR alone maintained higher accuracy than Snort alone, at 94.29% and 89.86% respectively.

Precision was also calculated for the three state conditions used in this experiment. Precision of recognition was highest when using both Snort and SoNSTAR together at 93.75%. SoNSTAR alone maintained higher precision again compared to Snort at 89.19% and 81.58% respectively.

The F-measure was highest when using both Snort and SoNSTAR together at 96.77%. SoNSTAR achieved a higher F-measure than Snort at 94.29% and 89.86% respectively.

NASA-Task Load Index results.

The NASA-Task Load Index results are shown in Table 6.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. NASA-Task Load Index results.

https://doi.org/10.1371/journal.pone.0195948.t006

Additional evaluation results.

Additional SoNSTAR evaluation results are shown in Table 7.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Additional SoNSTAR evaluation (index results).

https://doi.org/10.1371/journal.pone.0195948.t007

Table 8 shows participants’ opinions about whether using Snort and SoNSTAR alone or together would be best for monitoring.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 8. Additional SoNSTAR evaluation (preference results).

https://doi.org/10.1371/journal.pone.0195948.t008

Table 9 shows participants’ opinions about Snort and SoNSTAR from horrible (H) to fantastic (F).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 9. Horrible to fantastic evaluation.

https://doi.org/10.1371/journal.pone.0195948.t009

The most remarkable feedback was that a participant advised that if it is possible to add a visual panel showing the name, colour and image of what is generating the sound, this would help to distinguish the recorded sound in order to facilitate learning and confirmation. For example, some initially had trouble distinguishing between the sound of rain and the sound of rain on a roof, and a visual key might have helped to learn the sounds quicker.

Discussion

These experimental results clearly show improvements in monitoring when using sonification compared to the visual method only. Although the detection rate was 100% for the three state conditions, we can still see improvements in the accuracy, precision and F-measure scores for the sonification conditions. Although the training of participants was very brief and the computer security background of most of the participants was basic, they were able to use both systems well in a short time.

SoNSTAR can fill the gap between the network operator and the traffic environment by providing an auditory link between traffic behaviour and the operator’s mind. The experiment showed that users with a basic knowledge of computer network principles can use SoNSTAR to learn how flows are behaving inside network traffic and to recognise the typical combinations of the packet types within the flows. To ease the learning process, SoNSTAR generates log files holding counts of the packets within flows in addition to messages that indicate which flow event generated which specific sound. The IP-flow and traffic-flow log files together can explain in detail how the packet counts of the two flow types are related (see S3 Appendix IP flow log file and S4 Appendix Traffic flow log file in the supplemental material). These reports support the user in understanding the links between flag states, packet counts, flow counts, and the resultant sequences of sounds. The project repository [53] contains the source code and Max/MSP patches necessary to install and run SoNSTAR together with sample data files, example output stored in audio files, and instructions on how to use the system.

Some behaviours will create multiple sounds. For example, when performing a SYN scan, the attacker will send a number of packets with the SYN flag set to 1 to a number of targeted ports. If the port is open the receiver would send back a packet with the SYN flag set to 1 and the ACK flag set to 1 as a reply to accept the connection. The attacker either sends back a packet with the FIN flag set to 1 to cut the connection (the TCP half-open scan type) or sends two packets, the first with the ACK flag set to 1 to confirm the connection and then the second packet with the FIN flag set to 1 to cut the connection (the TCP connect scan type). If the port is closed the receiver would send back a packet with the RST flag set to 1, and if there is no response it means that the port is filtered.

As we set SoNSTAR to default settings, as soon as it receives many SYN in an IP flow, SoNSTAR will play the rain-on-a-roof sound and this would tell the user that an unusual number of SYN packets is arriving. If the TCP handshake was not correct, that event would generate a heavy rain sound which would tell the user that there is a problem with connecting to a specific IP address. If the number of SYN packets was high, SoNSTAR will play a thunder sound and this will tell the user that someone is scanning a large number of the system ports of a specific IP address. If the number was huge it would be considered a DoS attack and the sound of fire would be played. If the scanned system started to send out RST packets, SoNSTAR would play the sound of wind, confirming that it is a scan attack. This is a complex process, but SoNSTAR would deal with any changes in behaviour and play sets of sounds according to what events are happening in the network. The user could identify any new behaviour according to the set of sounds played.

Using such a tool to explore and tune a network is important due to the different natures of networks and the different expected behaviours with different thresholds. For example, this tool could be used to tune IDS settings to look for new features and events which could be used to identify threats on a particular network. Using SoNSTAR to draw a normal base line for a specific network behaviour would help to make rules and thresholds for specific environment which will raise the situation awareness in general.

Another advantage that SoNSTAR provides is that it generates log files which could help any user to learn and confirm how are packet types counts of IP flow effects recorded sound played and to evaluate theoretically any event and ideas of any new feature (Review the S3 Appendix IP flow log file in the supplemental material which contains a SYN scan). SoNSTAR could help users and network students to explore network protocols and to learn more about network traffic. The use of SoNSTAR would enable them to think directly about the logic of any behaviour in network traffic and would give them the opportunity to express their own ideas and to test and learn from them.

Using SoNSTAR reduced mental demand, temporal demand, effort and frustration rates compared to using Snort (visual tool) and this would be more obvious if the monitoring took place over long periods. This confirms that the use of SoNSTAR increases the situational awareness of the operator as it not only gives a warning of the attacks, but gives more than that so that the operator listens to what happens to the network directly allowing them to make links between behaviours and sounds. This relationship enables him or her to understand what is happening accurately by distinguishing the sounds and their meanings leading to comprehension of network state allowing decisions to be made directly without reference to any alarms reporting or going through long forensic processes to understand the behaviour.

Conclusion and further work

This study indicated that using sonification improved the monitoring process, even for people who have only basic knowledge of network monitoring. Using sound reduced the overall work load experienced by users. Participants were able to recognise and comprehend behaviours and decide which attack was performed which demonstrated it is possible for a human operator to quickly learn about the network environment in a way that results in increased situational awareness. Although the system could be evaluated manually by going through its log files, this experiment evaluated the practicality of using sonification in live monitoring tasks. The results suggest that using SoNSTAR to explore new event sand features would bring benefits for network monitoring in general and intrusion detection systems in particular.

There is more work that can be done to further improve attack detection and situational awareness. For example, using SoNSTAR’s feature extractor and combiner it was possible to define a number of traffic features. Discovering these from scratch is a time consuming task so public databases of features and feature combinations could be created. Developing a language for describing these discovered features and events based together with the log files to explain how these feature can be used could assist future designers of intrusion detection systems.

Vertical flow behaviour occurs when a single host receives many flows across a range of ports from a single source host. Horizontal flow behaviour occurs when a defined range of ports receives flows across a defined range of destination hosts. A major contribution of this research was to reduce the complexity of huge volumes of traffic in order to be comprehensively sonified by using IP flows in detecting network behaviour, especially vertical flow behaviours. The sonic mapping of network events based on packet type counts has not been seen before. We suggest further work to aimed at the identification of horizontal flow behaviours.

Other important future work consists of developing a method to create features for UDP, ICMP, IRC and other protocols for sonification by SoNSTAR so that larger quantities of common traffic can be covered. Also SoNSTAR has very high potential to represent SCADA (Supervisory Control and Data Aquisition) systems because of their unique nature. If their normal behaviour patterns are tested, zero-day vulnerability mitigation can be increased. Further study could be done to evaluate how SoNSTAR can be used to help students and computer professionals learn about network traffic.

Supporting information