The present invention relates to the protection of networks. In particular, the present invention relates to the protection of networks from the security risks associated with IoT devices.
The Internet of Things (IoT) has been defined as “the network of physical devices, vehicles, buildings and other items with embedded sensors and actuators”. Another definition of the IoT according to The IEEE can be found in their report titled “Towards a definition of the Internet of Things (IoT)”, revision 1 of which was published on 27 May 2015. Devices forming part of the IoT (otherwise referred to as IoT devices) are typically standard objects in which a computer system has been embedded (or attached) together with network connectivity and sensors and/or output devices (such as actuators) to enable data about the object to be collected and/or the object to be controlled. There has been a rapid growth in the popularity of IoT devices causing a significant increase in both the range and overall number of IoT devices that have been deployed into different networks. IoT devices can be found in a wide range of application areas, including, for example, farming, healthcare, infrastructure, logistics as well as in the home. For example, some of the kinds of IoT devices which can currently be found in home environments include smart lights, speakers, fans, environmental monitors, smoke detectors, doorbells, locks, burglar alarms and security cameras, as well as the personal IoT devices of the occupants, such as activity monitors, smart scales, blood pressure monitors, and so on. The network connectivity for these devices is typically provided by connecting them to a local network, such as the wireless network provided by a home router (either directly or indirectly through another device connected to the local network). However, other methods of providing network connectivity for IoT devices are known, including, for example, by providing connectivity through a cellular network.
The computer systems that are embedded in (or attached to) an object to create an IoT device are typically low-powered as a result of physical and/or financial limitations. This in turn places constraints on the computational ability of the embedded (or attached) computer system. Due to these computational constraints, IoT devices are typically controlled by an application-specific operating system which is tailored toward the specific functions of the IoT device. However, because of this, IoT devices can present a security risk to any network that they are connected to. In particular, due to the application-specific nature of the operating system, it is generally not possible to use traditional security measures, such as anti-malware and firewall systems that can be applied to user endpoints such as computer, tablets and smart phones. Additionally, due to the customized nature of the application-specific operating system and other software running on an IoT device, the likelihood of exploitable vulnerabilities being present in IoT devices may be higher. Furthermore, due to the wide range of different types of models of IoT devices that may be present in a network, as well as the range of different manufacturers responsible for maintaining the application-specific operating system and software of these devices, updating the IoT devices to mitigate known security vulnerabilities can be difficult, time consuming and typically lags behind updates applied to more conventional computer systems, such as laptops, tablets and mobile phones, especially in home networks.
Looking to the future, it is expected that both the range and overall number of IoT devices will continue to increase. This, combined with security weaknesses of IoT devices, make IoT devices an increasingly attractive target for attackers to seek to exploit either directly or via malware. Even if gaining the ability to retrieve data from and/or control the IoT device itself is not of particular interest to an attacker (which may not necessarily be the case), the exploitation of vulnerabilities which may be more likely to be present in such devices can provide useful starting points from which to launch further attacks either inside or outside of the network to which the IoT device is connected. As an example, an attacker may be able to exploit a vulnerable IoT device to gather information passing through the network from other devices on the network, including from user computing devices such as laptops, tablets and smart phones. As a further example, an attacker may look to exploit (very) large numbers of vulnerable IoT devices from across a large number of networks in order to launch a Distributed Denial-of-Service (DDoS) attack on a computer system accessible to those IoT devices via the internet. This can cause problems not only for the targeted computer system, but also for the networks over which the DDoS attack traffic is carried due to the increased amount of traffic such attacks can generate.
Accordingly, it would be beneficial to mitigate these disadvantages.
In a first aspect, there is provided a computer implemented method of protecting a network, the method comprising: gathering traffic data for the network; identifying a set of IoT devices in the network based on the output from a machine learning model for classifying IoT devices using features extracted from the traffic data that are indicative of an IoT device; and causing one or more predetermined actions to be taken in respect of the set of IoT devices to protect the network.
Through the use of the machine learning model to classify IoT devices, the present invention enables a set of IoT devices in a network to be identified automatically from traffic data for the network, thereby enabling action to be taken to protect the network from those IoT devices without requiring any interaction from an administrator of the network. Furthermore, as a result, the threat posed by those IoT devices to computer systems and other networks outside of the network (such as an Internet Service Provider's network) may also be reduced.
The method may be performed by a router device for the network. The traffic data may comprise data gathered from logs maintained by the router device for network services which are provided by the router device. The traffic data may further comprise data gathered from other computing devices of the network.
The traffic data may comprise indications of one or more, or all, of: successful network connections; network connection attempts; network connection terminations; packet filtering operations; network address translation operations; port translation operations; network session operations; layer 2 connections; access control operations; authentication operations; router advertisements; network boot operations; DNS requests and responses; and HTTP requests and responses.
The method may further comprise extracting one or more features that are indicative of an IoT device from the traffic data. The one or more features may relate to operational parameters of the network traffic, such as one or more, or all, of: the number of packets of data transmitted by each device; the number of packets of data received by each device; the frequency with which traffic is transmitted by each device; the frequency with which traffic is received by each device; the average size of packets of data which are transmitted by each device; the average size of packets of data which are received by each device; the variance in the sizes of packets which are transmitted by each device; the variance in the sizes of packets which are received by each device; the ratio of traffic-in against traffic-out for each device; the number of endpoints with which each device communicates; the average duration of communication sessions for each device; the typical times when communication sessions occur for each device; and/or the times and duration that each device is online.
The method may further comprise generating the machine learning model using an unsupervised machine learning algorithm and a training set of data obtained from previously gathered traffic data for the network. The machine learning algorithm may comprise a shallow learning algorithm.
The method may further comprise communicating with another computer system to identify the set of IoT devices. The method may receive an indication from the other computer system of one or more features that are indicative of an IoT device to be extracted from the traffic data, extract the one or more features from the traffic data, provide the one or more features to the other computer system and receive an indication of the set of IoT devices from the other computer system. The method may generate a profile of the computational abilities of the routing device and determine that processing to identify IoT devices in the network is to be performed using the processing resources provided by the other computer system based, at least in part, on the profile.
The one or more predetermined actions may comprise placing the identified set of IoT devices into a separate VLAN. The one or more predetermined actions may comprise performing targeted patching of the identified set of IoT devices. The one or more predetermined actions may comprise comparing the set of IoT devices to a previously identified set of IoT devices and raising an alarm if there are any differences.
In a second aspect, there is provided a computer-implemented method for protecting a network comprising: obtaining a machine learning model for identifying IoT devices in a network using features extracted from traffic data for that network that are indicative of an IoT device; providing an indication to a routing device associated with a network of one or more features that are indicative of an IoT device to be extracted from the traffic data; receiving the one or more features from the routing device; identifying a set of IoT devices in the network using the machine learning model and the one or more received features; and providing an indication of the set of IoT devices to the routing device.
The machine learning model may be learnt from training data that is based on the traffic data from a plurality of networks. The machine learning algorithm may comprise an unsupervised learning algorithm. The machine learning algorithm may comprise a deep learning algorithm.
The one or more features relate to operational parameters of the network traffic, such as one or more, or all, of: the number of packets of data transmitted by each device; the number of packets of data received by each device; the frequency with which traffic is transmitted by each device; the frequency with which traffic is received by each device; the average size of packets of data which are transmitted by each device; the average size of packets of data which are received by each device; the variance in the sizes of packets which are transmitted by each device; the variance in the sizes of packets which are received by each device; the ratio of traffic-in against traffic-out for each device; the number of endpoints with which each device communicates; the average duration of communication sessions for each device; the typical times when communication sessions occur for each device; and/or the times and duration that each device is online.
In a third aspect, there is provided a computer system for protecting a network comprising a processor and a memory storing computer program code which, when executed by the processor cause the processor to perform a method according to the first or second aspects. The computer system may be arranged to function as a router device for the network.
In a fourth aspect, there is provided a system for protecting a network which comprises: a plurality of router devices, each router device being associated with a respective network and being configured to perform a method according to the first aspect to protect that network; and a computer system configured to perform a method according to the second aspect.
In a fifth aspect, there is provided a computer program which, when executed by one or more processors, is arranged to cause the processor to carry out the method set out above.
Embodiments of the present invention will now be described by way of example only, with reference to the accompanying drawings, in which:
The storage 110 can be any volatile read/write storage device such as a random access memory (RAM) or a non-volatile storage device such as a hard disk drive, magnetic disc, optical disc, ROM and so on. The storage 110 can be formed as a hierarchy of a plurality of different storage devices, including both volatile and non-volatile storage devices, with the different storage devices in the hierarchy providing differing capacities and response times, as is well known in the art.
The processor 120 may be any processing unit, such as central processing unit (CPU) which is suitable for executing one or more computer programs (or software or instructions or code). These computer programs may be stored in the storage 110. During operation of the system 100, the computer programs may be provided from the storage 110 to the processor 120 via the one or more buses 140 for execution. One or more of the stored computer programs are computer programs which, when executed by the processor 120, cause the processor 120 to carry out a method according to an embodiment of the invention (and so configure the system 100 to be a system 100 according to an embodiment of the invention).
The one or more input/output (I/O) interfaces 130 provide interfaces to devices 150 for the input or output of data, or for both the input and output of data. The one or more interfaces 130 may include one or more user input interfaces 130a for connecting to devices which can receive input from a user of the system 100, such as a keyboard 150a or mouse 150b. The one or more interfaces 130 may include one or more user output interfaces 130b which can provide an output to the user of the system 100, such as a display 150c or monitor. In some cases a single device, such as a touch screen display, may be connected to both an input interface 130a and an output interface 130b and used both to receive input from the user of the system 100 and provide output to the user of the system 100. The one or more interfaces 130 may include one or more network interfaces 130c which enable the computer system to communicate with other computer systems via one or more networks 160. Other interfaces (not shown) may also be present in the computer system. Indeed, there are many other types of devices (not shown) which may be used with system 100 as is well known in the art. For example, the system 100 can include interfaces to various sensors and actuators which enable it to monitor and/or interact with its environment.
It will be appreciated that the architecture of the system 100 illustrated in
The router device 210 manages the flow of network traffic between the computer systems 220 within the network 200. In some embodiments, the router device 210 enables the computer systems 220 within the network 200 to communicate with other networks (not shown in
The computer systems 220 are communicatively coupled to the router device 210 via any suitable data links. In some embodiments, the router device 210 is communicatively coupled to the computer systems 220 via data links established over cable media, such as wires or optic cables. In some embodiments, the router device 210 is communicatively coupled to the computer systems 220 via data links established over wireless media, such as via WiFi, LiFi or Cellular communication. In some embodiments, the router device 210 is communicatively coupled to the computer systems 220 via data links established over more than one media, with each computer system 220 using a respective media to communicate with the router device 210. In some embodiments, some of the computer systems 220 may be connected via a mesh network, such computer systems 220 thereby communicating with the router device indirectly via another one of the computer systems 220 in the mesh network with a direct communication link to the router device 210.
The computer systems 220 in the network 200 may be any type of computer system 100 as described above in conjunction with
In an embodiment, the method 300 is performed by a router device for the network 200, such as router device 210. Since router devices are present in almost all networks, the use of a router device 210 to perform the method 300 allows the method 300 to be performed without requiring additional devices to be installed in the network 200. This can reduce the costs of implementing the invention, as well as reducing the complexity and time required to set up a network which is protected by method 300. Additionally, the router device 210 may be better placed to take action to mitigate the risks posed by the IoT devices (due to its role of managing the traffic in the network) thereby simplifying the implementation of the method 300. Of course, in other embodiments, the method 300 may be performed by another computer system 220 within the network 200, either as a separate device dedicated to protecting the network 200 using method 300 or by any other computer system 220 in the network 200 which may perform other functions in addition to the method 300.
At an operation 310, the method 300 gathers traffic data from the network 200.
The traffic data represents the traffic (or data flows or communication) that is occurring in the network 200. That is to say, the traffic data represents the traffic flowing to each computer system 220, from each computer system 220 or both. The traffic data may include traffic flowing between computer systems 220 within the network 200 as well as traffic flowing between computer systems 220 within the network 200 and those outside the network (in embodiments where the router 210 enables the computer systems 220 within the network 200 to communicate with other networks).
Where the method 300 is performed by a router device 210 for the network 200, at least some of the traffic data that is gathered may be gathered from logs that are maintained by the router device for network services which are provided by the router device. As will be appreciated, router devices 210 commonly provide additional network services that may be used by the network 200 to which they are connected, especially those router devices 210 that are intended for use in a home environment. As examples, such router devices 210 may also provide firewall and DHCP services to the network 200 (although any type and combination of additional network services may be used and will typically be different for different router devices). The logs for such network services typically include data which represents the traffic occurring in the network 200 and as such are a good source of traffic data for the network 200. Examples of the types of logs that may be available on a typical router device 210 include:
It will of course be appreciated that the names of the logs and the type of traffic data they provide may vary depending on the software used in a particular router and that fewer, additional and/or different logs may be used in other embodiments to gather traffic data for the network from logs within the router device 210 accordingly.
Additionally, in some embodiments the router device 300 may gather traffic data from other computing devices 220 within the network 200. In particular, other computer systems 220 may provide network services for the network 200 and can provide data representing the traffic occurring in the network 200. For example, a separate computer system 220 may be used to provide a firewall service for the network 200, in which case, logs may be retrieved from that system 220 and used as traffic data for the network 200.
In a similar manner to that discussed above, where the method 300 is performed by a device 220 other than the router device 210, the traffic data may be gathered from logs maintained locally on that device (for example, from logs relating to network services provided by that device 220). Additionally or alternatively, the traffic data may be gathered from other computer systems 220 on the network 200.
Regardless of where the traffic data is collected from, in embodiments the traffic data comprises any combination of one or more, or all, of: successful network connections; network connection attempts; network connection terminations; packet filtering operations; network address translation operations; port translation operations; network session operations; layer 2 connections; access control operations; authentication operations; router advertisements; network boot operations; DNS requests and responses; and/or HTTP requests and responses. Although, of course, in other embodiments additional and/or different types of traffic data may be gathered.
Having gathered traffic data from the network 200 at operation 310, the method 300 proceeds to an operation 320.
At operation 320, the method 300 uses the output from a machine learning model to identify a set of IoT devices 240 in the network 200.
The machine learning model is a model that is the outcome from the training of a machine learning algorithm on a training set (or test set) of traffic data. This training process produces a trained model (the machine learning model), which is able to classify a computer system 220 as being either an IoT device 240 or a non-IoT device 230 based on one or more features extracted from the traffic data. Due to the fact that IoT devices typically operate largely autonomously and are application-specific, their behaviour tends to differ from that of other computer systems. There are therefore numerous features that may be extracted from traffic data which are likely to distinguish an IoT device 240 from a non-IoT device 230. As examples, the regularity (or frequency) with which traffic is transmitted (and/or received), the average size and/or variance in sizes of packets of data which are transmitted (and/or received), the average duration of communication sessions, the typical times and the time period over which communication sessions occur, the number of endpoints with which communication takes place, the ratio of traffic-in against traffic-out, source and destination IP addresses, numbers of packets, timestamps, device online durations, text and so on may all differ in distinct ways for IoT devices as compared to non-IoT devices 230.
Those skilled in the art will appreciate that these features of the traffic data are merely provided as examples and that other features of the traffic data which can distinguish IoT devices 240 from non-IoT devices 230 may be used instead (or in addition). Whilst the characteristics of specific IoT devices 240, such as a particular model of IoT device made by a particular manufacturer, are likely to differ in some ways from those of other IoT devices 240, it will be appreciated that in other aspects there are commonalities which are shared between large numbers of IoT devices. It is therefore possible for a machine learning model to be trained to identify IoT devices generally based on these underlying common features which are shared between large numbers of IoT devices and which are distinct from the characteristics of non-IoT devices. By using a machine learning model trained to distinguish between IoT devices and non-IoT devices using such features at a general level (rather than identifying specific models of IoT device), the method 200 is more adaptable and is more likely to be able to continue to operate effectively when faced with new kinds of IoT devices which were not included in the training set upon which the machine learning model was trained.
As will be known by those skilled in the art, two types of machine learning algorithms are supervised learning algorithms and unsupervised learning algorithms. Supervised learning algorithms require training to be performed using a set of labeled training (or test) data. That is to say, for each training data input, the algorithm needs to know whether that training data input represents an IoT device or a non-IoT device (in the context of this invention). Examples of supervised learning algorithms include decision trees, random forests, k nearest neighbour, linear support vector classifiers (SVC), logistic regression, naïve Bayes, neural networks and support vector regression (SVR). Unsupervised learning algorithms do not require the data that they are trained on to be labeled. Examples of unsupervised learning algorithms include k-means clustering, n nearest neighbour, dimensionality reduction, neural networks, principal component analysis and singular value decomposition and support vector machines. Whilst embodiments of the invention may make use of any suitable supervised or unsupervised learning algorithm, as known to those skilled in the art, preferably unsupervised learning algorithms are used to create the machine learning model. In particular, given the large numbers of different types of IoT devices, creating a labeled training set of data to cover those IoT devices, as required for supervised learning algorithms, is laborious, time consuming and costly. Additionally, supervised learning algorithms are more prone to overfitting the model that they produce to the specific types of IoT devices represented in the training set of data. This means that the model that is produced by supervised learning algorithms may be less adaptable. That is to say, the model may be less likely to detect new types of IoT devices which were not represented in the training set of data. Therefore, given the rate at which new types of IoT devices are being created an unsupervised learning algorithm is preferred as such algorithms may be more adaptable and are, in any case, easier to retain to account for new types of IoT devices as they do not require labeled training data to be obtained.
There are a variety of different ways in which the output from the machine learning model may be obtained by operation 320. These are discussed further below in conjunction with the discussion of the embodiments illustrated in
At operation 330, the method 300 causes one or more predetermined actions to be taken in respect of the set of IoT devices to protect the network.
As discussed above, IoT devices typically are more likely to present a security risk to any network that they are connected to. Therefore, by identifying a set of IoT devices 240 in the network at operation 320, the method 300 can take various measures to mitigate the risk presented by these devices and, in so doing, protect the network 200 from such risks.
In some embodiments, the network 200 is protected from the risks presented by the IoT devices 240 by placing the identified set of IoT devices 240 into a separate Virtual Local Area Network (VLAN). That is to say, the predetermined actions may comprise placing the identified set of IoT devices into a VLAN which is separate from any non-IoT devices 230 in the network 200. Any devices which have not yet been classified can be treated in any appropriate manner. For example, a permissive-based approach can assume that unclassified devices are non-IoT devices and allow them to be in the same VLANs as other non-IoT devices. However, a more restrictive approach can place them the unclassified devices into their own “unclassified” VLAN or else place them in the same VLAN as the classified IoT devices.
As will be understood by those skilled in the art, although different VLANs exist on the same underlying physical network (or LAN) they act as though they are completely separate networks. This means that the traffic data on each VLAN is separated (or segregated) from the traffic data on any other VLAN. The router device 210 can be configured to apply different rules to the devices in different VLANs. For example, the devices in one VLAN, such as a VLAN containing the non-IoT devices 230, may be allowed to initiate connections to devices in another VLAN, such as a VLAN containing the IoT devices 240. Meanwhile, devices in the other VLAN containing the IoT devices 240 may not be allowed to initiate connections to devices in the VLAN containing the non-IoT devices 230. Similarly, where the router device 210 enables computer systems 220 in the network 200 to communicate with another network (or a wider network, or the internet), different rules can be applied to the devices 220 in one VLAN from those in another VLAN. For example, devices in a VLAN containing non-IoT devices 230 may be allowed to make any number of connections to any number of computer systems in the other network. Meanwhile, devices in a VLAN containing IoT devices 240 may be more restricted in the connections that they are permitted to make. Therefore, by placing the IoT devices 240 into a separate VLAN from the non-IoT devices 230, the traffic of the non-IoT devices 230 is not accessible to the IoT devices 240. This eliminates (or at least reduces) the risk of an attacker being able to exploit a vulnerable IoT device to gather information passing through the network from the non-IoT devices on the network. Furthermore, due to the restrictions that may be imposed by the router device 210 on the communications of the devices 220 in the VLAN containing IoT devices 240, the possibility for malware to spread or for attackers to launch attacks (such as DDoS attacks) on other computer systems, either inside or outside of the network 200, may be reduced.
In some embodiments, the predetermined actions that are taken in respect of the identified set of IoT devices 240 includes performing targeted patching of the identified devices. In some embodiments, any IoT devices 240 that are identified as requiring a patch to address a vulnerability may be quarantined prior to the patch being applied (for example, by placing them into a separate VLAN until the patch is applied).
In some embodiments, the predetermined actions that are taken in respect of the identified set of IoT devices 240 includes comparing the set of IoT devices to a previously identified set of IoT devices and raising an alarm (or notification or alert) if there are any differences. For example, where the method 300 is performed periodically, the previous set of IoT devices may be the set of IoT devices identified in the previous iteration of the method 300. In some embodiments, the composition of the set of IoT devices 240 may be considered to have changed compared to the previously identified set of IoT devices if a new device which was either not previously in the network, or which was not previously classified as being an IoT device (for example, due to using an older version of a machine learning model), has been classified as being an IoT device. Additionally or alternatively, the composition of the set of IoT devices 240 may be considered to have changed compared to the previously identified set of IoT devices if a device which was previously classified as being an IoT device is either no longer present in the network or is no longer classified as being an IoT device. In some embodiments, the alarm which is raised identifies the device(s) which have changed (i.e. the alarm identifies the new and/or removed IoT devices).
Having caused one or more predetermined actions to be taken in respect of the set of IoT devices to protect the network at operation 330, the method 300 proceeds to operation 340.
At operation 340, the method 300 determines whether it should repeat operations 310-330. That is to say, in some embodiments, the method 300 is performed iteratively whilst in others it is not. By performing the method 300 iteratively, the network 200 can be periodically (or sporadically) monitored by waiting until an appropriate time to repeat the method. For example, operations 310-330 could be repeated at regular intervals, such as every minute, hour, day, week and so on (or indeed any time period in between). Alternatively the method 300 may be performed in response to another event. For example, when a new computer system 220 is detected in the network 200 (such as, for example, when a new DHCP lease is requested). Similarly, the two approaches may be combined in some embodiments to allow both regular and responsive monitoring of the network. By performing the method 300 iteratively in any of these ways, the method 300 can detect any additional IoT devices 240 that have been connected to (or removed from) the network 200 since the previous iteration. Additionally, in embodiments where the machine learning model may be updated (or re-trained) based on subsequently collected traffic data (as discussed further below), subsequent iterations of the method 300 may use the updated model to identify any IoT devices which may have been incorrectly classified as non-IoT devices by the previous model (or indeed to identify any non-IoT devices which may have been incorrectly classified as IoT devices). Where the method is performed iteratively, when it is determined that operations 310-330 should be repeated, the method 300 re-iterates back to operation 310. Of course, in other embodiments, the method may not be performed iteratively, in which case the method 300 ends.
Through the use of a machine learning model to identify a set of IoT devices in the network based on traffic data, the method 300 enables action to be taken to protect a network 200 from the risks associated with IoT devices 240, without requiring explicit configuration as to which devices are the IoT devices and without requiring any modifications being made to the devices 220 operating in the network 200.
As discussed above, there are a variety of different ways in which the output from the machine learning model may be obtained by operation 320 of method 300. These will now be discussed further below in conjunction with the discussion of the embodiments illustrated in
As for the method 300 described above in relation to
At operation 410, the method 400 determines whether the machine learning model needs to be learned (or re-learned).
If the model needs to be learned (or re-learned), the method 400 proceeds to an operation 420. This may be the case, for example, during an initial run of method 400 on a particular computer system 100, such as router device 220, whereby no model that was learnt by a previous run of method 400 may be available for use. Similarly, it may be determined that a model that is available for use, for example from a previous run of method 400, is sufficiently old that it should be updated (i.e. re-learned).
At operation 420, the method 400 generates a machine learning model using a machine learning algorithm and a training set of data derived from the set of data gathered at operation 310. The machine learning model is generated using an unsupervised learning algorithm such as any of those listed above. In order to train the unsupervised learning algorithm a set of features is extracted from the training set of data and used as input from the model. In some embodiments, the features that are extracted are predetermined. In other embodiments the features are determined through the use of feature learning techniques (otherwise referred to as automated feature engineering, learning feature engineering or deep feature synthesis). In such embodiments, any suitable method of feature learning known in the art may be used as will be apparent to those skilled in the art including, for example, the use of clustering methods and/or principal component analysis and/or deep learning using recurrent neural networks. The extraction of the features from the traffic data creates a locally processable training set, for example by representing various features of the traffic data as numerical values which may be normalized or scaled into useful ranges (such as an interval from −1 to 1) for more efficient processing (in terms of hardware/network resources).
In some embodiments, the machine learning algorithm is a so-called shallow learning algorithm (also referred to as a lightweight learning algorithm). Due to the lower computational power required to train a shallow learning algorithm (compared to deep learning algorithms) the use of a shallow learning algorithm enables the method 400 to be performed on lower-powered computer systems, such as some router devices (especially those typically found in a home network environment), which have less computational, storage and/or I/O resources available for performing the algorithm (and may not therefore be able to perform a more deep or heavyweight learning algorithm). Examples of shallow learning algorithms which may be used include logistic or linear regression or support vector machines, although any suitable light-weight machine learning algorithm may be used.
Having generated a machine learning model using a training set of data derived from the set of data gathered at operation 310, the method 400 returns to operation 310 to gather more traffic data upon which to operate the model.
If, at operation 410, it is determined that the model does not need to be learned (or re-learned), the method 400 proceeds to an operation 430. This may be the case, for example, where a machine learning model that was learnt during a previous run of method 400 is available for use and is considered to be sufficiently current that it does not need to be updated. In some embodiments, the machine learning model that is to be used is not generated by the computer system 100 which is performing the method 400 (such as router device 220). For example, the machine learning model may be learnt on a separate computer system and supplied pre-stored in a storage 110 of the computer system 100 that is to run the method 400. In such embodiments, the machine learning model is simply retrieved from the storage 110.
At operation 430, the method 400 extracts one or more features that are indicative of an IoT device from the traffic data as required for the model. As will be understood by those skilled in the art, the set of features that are extracted in order to use the model at operation 430 may be different from the features that were extracted in order to train the model during operation 420. In particular, during training of the model, various features may be discarded as being less suitable for distinguishing between the devices in the network. The features may relate to operational parameters of the network traffic. That is to say, they may relate to the properties of the network traffic, rather than its content. As examples, the types of features that may be extracted at operation include the regularity (or frequency) with which traffic is transmitted (and/or received) from each computer system 220 in the network, the average size and/or variance in sizes of packets of data which are transmitted (and/or received) from each computer system 220 in the network, the average duration of communication sessions from each computer system 220 in the network, the typical times and the time period over which communication sessions occur from each computer system 220 in the network, the number of endpoints with which communication takes place for each computer system 220 in the network, the ratio of traffic in against traffic out for each computer system 220. However, those skilled in the art would understand that these are merely exemplary and that different combinations and different features may be used in other embodiments. Having extracted the features needed to use the model, the method 400 proceeds to an operation 440.
At operation 440, the method 400 inputs the features extracted at operation 430 into the machine learning model. The output from using the machine learning model with the extracted features is a classification, for each of the computer systems 220 for which features were extracted from the traffic data, as to whether that computer system 220 is an IoT device or not. Using this output, at operation 440, the method identifies a set of IoT devices in the network.
As for the method 300 described above in relation to
As for the method 300 described above in relation to
As for the method 300 described above in relation to
At optional operation 610, the method 600 generates a profile of the computational abilities of the system 220 that is performing the method 600, such as routing device 210. The profile includes various metrics of the system, including for example, regarding its processing, memory, I/O and throughput capabilities. In some embodiments, this profile is generated by actively probing (or testing) the capabilities of the system 220. In other embodiments, the profile is generated by retrieving stored data indicating the system's capabilities. In yet further embodiments, a combination of both probing and retrieval of stored data are employed. Having generated the profile, the method 600 proceeds to operation 620.
At operation 620, the method 600 determines whether remote processing (or server-side or cloud-side) resources should be used to assist in the identification of IoT devices 240 in the network 200. As discussed above, the remote processing resources are provided by another computer system, different from the system performing the method 600. In some embodiments the other computer system is a server 520 or cloud service 530 accessible to the system performing the method 600 via another network 510, such as the internet.
In embodiments where optional operation 610 is performed to generate a profile of the computational abilities of the system 220 that is performing the method 600, the determination as to whether remote processing resources should be used, may be based on the profile. That is to say, in some embodiments, it is determined at operation 610 whether the device 220 that is performing the method 600 is able to classify the devices 220 in the network 200 solely using local (or edge-side) resources and if not, determine that remote processing resources should be used. As an example, in some embodiments, the profile comprises a score indicating the computational abilities for the device 220 that is performing the method 600 and the score is compared to a predetermined threshold to determine whether remote processing should be used. Alternatively or additionally, in such embodiments, the operation 610 may use the profile to assess whether there is sufficient networking capacity to be able to upload the amount of data required to utilize remote processing resources and, if not, determine that remote processing resources should not be used.
Naturally, in these and other embodiments, a locally stored value (i.e. indicating a user preference or device setting) may be used as part of the determination to use remote processing resources. For example, if a user preference is stored specifying that remote processing resources should not be used, the method 600 at operation 620 can determine that remote processing resources should not be used.
If it is determined, at operation 620, that remote processing resources should not be used, in some embodiments the method 600 may instead, at an optional operation 630 attempt to use local processing resources to identify the set of IoT devices by switching to use the method 400 described in conjunction with
If it is determined, at operation 620, that remote processing resources should be used, the method 600 proceeds to an operation 640.
At operation 640, the method 600 communicates with the other computer system, such as server 520, to identify a set of IoT devices 240 in the network 200. There are a variety of different ways in which the set of IoT devices 240 may be identified by communicating with the other computer system. These are discussed further below in conjunction with the discussion of the embodiments illustrated in
As for the method 300 described above in relation to
As for the method 300 described above in relation to
At an operation 710, the method 700 obtains a machine learning model for identifying IoT devices in the network. In some embodiments, the machine learning model is simply retrieved from a storage or network location. That is to say, a machine learning model that has already been trained may be provided to the computer system performing the method 700. In other embodiments, as discussed in more detail in conjunction with
At operation 720, the method 700 communicates with a routing device for the network to identify a set of IoT devices from traffic data gathered from the network by the routing device. As for the operation 640 performed by the method 600 described in conjunction with
It will be appreciated that, in some embodiments, the operations 710 and 720 of method 700 are largely decoupled from each other. For example, in some embodiments, operation 720 may be performed in response to the initiation of communication by a remote device, such as the router 210, to collaboratively identify IoT devices 240 in the network 200 and may use the model that has been retrieved by operation 710 at some point prior to the initiation of communication by the remote device. Similarly, in some embodiments, operation 720 of method 700 communicates with multiple different routing devices 210, potentially substantially in parallel, in order to collaboratively identify respective sets of IoT devices in each network.
At an operation 810, the method 800 receives traffic data for one or more networks. For example, the traffic data may be provided by respective router devices 210 operating in the one or more networks. The networks from which the traffic data is received need not necessarily be networks in which embodiments of the invention are operating to protect the network. Of course, in some embodiments, the traffic data may be exclusively received from networks in which embodiments of the invention are operating to protect the network. In other embodiments, the traffic data may be exclusively received from networks in which embodiments of the invention are not operating to protect the network. In yet other embodiments, the traffic data may be received from both types of network.
At an operation 820, the method 800 generates a machine learning model using a machine learning algorithm and a training set of data obtained from the received traffic data. Whilst any type of unsupervised machine learning algorithm may be used. In some preferred embodiments, the machine learning algorithm is a heavy weight algorithm, such as a deep learning or text classification. Such heavy weight machine learning algorithms may take full advantage of the additional resources typically available on the other computer system to achieve a high level of classification accuracy for the generated model.
Again, it will be appreciated that the operations 810 and 820 may be performed substantially independently from each other. In particular, the method 800 may continually (or periodically or sporadically) receive traffic data for one or more networks and may periodically (or sporadically) train or re-train the machine learning model based on a training set of data derived from the traffic data that has been received up until that point. The timing (and frequency) for the training of the machine learning model can be completely independent of the receipt of the traffic data.
Although not illustrated in
At a step 910, the operation 640 (performed, for example, by router device 210) provides the traffic data that was gathered at operation 310 to the other computer system. This traffic data is received by the other computer system at a step 1010 of operation 720. In some embodiments, the traffic data that is received may be stored for use in training or re-training machine learning models in the future (i.e. the data used in operation 810 of the method illustrated by
At a step 1020, the operation 720 performed by the other computer system extracts one or more features that are indicative of an IoT device from the traffic data.
At a step 1030, the operation 720 performed by the other computer system uses the machine learning model obtained in operation 710 to identify IoT devices 240 in the network 200 based on the extracted features.
At a step 1040, the operation 720 performed by the other computer system provides the indication of the set of IoT devices back to the device 220, such as routing device 210. This indication is received at a step 920 of the operation 640 (performed, for example, by router device 210).
By providing the traffic data to the other computer system to process, the steps illustrated in
At a step 1210, the operation 720 performed by the other computer system provides an indication to the routing device of one or more features that are indicative of an IoT device. These features are the features that are required as inputs to the machine learning model obtained in operation 710. This indication is received by at a step 1110 of operation 640 (performed, for example, by router device 210).
At a step 1120, the operation 640 (performed, for example, by router device 210) extracts the features, as indicated by the other computer system, from the traffic data.
At a step 1130, the operation 640 provides the extracted features to the other computer system, which receives them at a step 1220 of operation 720.
At a step 1230, the operation 720 performed by the other computer system uses the machine learning model and the received features to identify the set of IoT devices.
At a step 1240, the operation 720 performed by the other computer system provides an indication of the set of IoT devices to the device 220, such as router device 210, in the network. This is received at step 1140 of operation 640.
By extracting the features on the device 220 within the network 200 the steps illustrated in
Although in this description of the invention, the machine learning model that is used is described as classifying a device as being an IoT device (or not), it will be appreciated that IoT devices may be further subcategorised. For example, some IoT devices are entirely concerned with the provision of information from sensors embedded in the object and/or its environment. Other IoT devices are entirely concerned with controlling the object and/or its environment using embedded actuators. Yet other IoT devices may provide a combination of both. Some IoT devices enable a user to interact with the device from a conventional computer system, such as a laptop, tablet or mobile phone to receive the sensed data or control the object. Alternatively, some IoT devices only communicate with other computer systems, such as other IoT devices or cloud services to provide data about the object or receive input for controlling the object. Yet other IoT devices sense user input through interaction with the object (such as by pushing buttons on the object) and use this input to control the object itself and/or another IoT device. Each of these different type (or subcategory) of IoT device is likely to exhibit its own unique set of characteristics and therefore be distinguishable from other types (or subcategories) of IoT devices (in addition to be distinguishable from non-IoT devices). Therefore, whilst in some embodiments the machine learning model is generally used to classify a device as being an IoT device (or not), in other embodiments, the machine learning model may classify a device as being a specific type (or subcategory) of IoT device (or not). In yet further embodiments, multiple machine learning models may be used, each tailored to one or more respective subcategories of IoT device, to classify the devices as being an IoT device belonging to one or more subcategory of IoT device (or not). In such embodiments, the predetermined actions that are taken may vary between the different subcategories of IoT device to account for the specific threats that are most likely to impact such devices. In some such embodiments, separate VLANs may be maintained for each subcategory of IoT device.
Insofar as embodiments of the invention described are implementable, at least in part, using a software-controlled programmable processing device, such as a microprocessor, digital signal processor or other processing device, data processing apparatus or system, it will be appreciated that a computer program for configuring a programmable device, apparatus or system to implement the foregoing described methods is envisaged as an aspect of the present invention. The computer program may be embodied as source code or undergo compilation for implementation on a processing device, apparatus or system or may be embodied as object code, for example. Suitably, the computer program is stored on a carrier medium in machine or device readable form, for example in solid-state memory, magnetic memory such as disk or tape, optically or magneto-optically readable memory such as compact disk or digital versatile disk etc., and the processing device utilizes the program or a part thereof to configure it for operation. The computer program may be supplied from a remote source embodied in a communications medium such as an electronic signal, radio frequency carrier wave or optical carrier wave. Such carrier media are also envisaged as aspects of the present invention. It will be understood by those skilled in the art that, although the present invention has been described in relation to the above described example embodiments, the invention is not limited thereto and that there are many possible variations and modifications which fall within the scope of the invention. The scope of the present invention includes any novel features or combination of features disclosed herein. The applicant hereby gives notice that new claims may be formulated to such features or combination of features during prosecution of this application or of any such further applications derived therefrom. In particular, with reference to the appended claims, features from dependent claims may be combined with those of the independent claims and features from respective independent claims may be combined in any appropriate manner and not merely in the specific combinations enumerated in the claims.
Number | Date | Country | Kind |
---|---|---|---|
19161087.2 | Mar 2019 | EP | regional |
19161094.8 | Mar 2019 | EP | regional |
Filing Document | Filing Date | Country | Kind |
---|---|---|---|
PCT/EP2020/055502 | 3/3/2020 | WO | 00 |