The present disclosure relates generally to computer networks, and, more particularly, to optimizing a voting process used by classifiers distributed within a network.
Low power and Lossy Networks (LLNs), e.g., sensor networks, have a myriad of applications, such as Smart Grid and Smart Cities. Various challenges are presented with LLNs, such as lossy links, low bandwidth, battery operation, low memory and/or processing capability of a device, etc. Changing environmental conditions may also affect device communications. For example, physical obstructions (e.g., changes in the foliage density of nearby trees, the opening and closing of doors, etc.), changes in interference (e.g., from other wireless networks or devices), propagation characteristics of the media (e.g., temperature or humidity changes, etc.), and the like also present unique challenges to LLNs.
One type of network attack that is of particular concern in the context of LLNs is a Denial of Service (DoS) attack. Typically, DoS attacks operate by attempting to exhaust the available resources of a service (e.g., bandwidth, memory, etc.), thereby preventing legitimate traffic from using the resource. A DoS attack may also be distributed, to conceal the presence of the attack. For example, a distributed DoS (DDoS) attack may involve multiple attackers sending malicious requests, making it more difficult to distinguish when an attack is underway.
The embodiments herein may be better understood by referring to the following description in conjunction with the accompanying drawings in which like reference numerals indicate identically or functionally similar elements, of which:
According to one or more embodiments of the disclosure, voting optimization requests that identify a validation data set are sent to a plurality of network nodes. Voting optimization data is received from the plurality of network nodes that was generated by executing classifiers using the validation data set. A set of one or more voting classifiers is then selected from among the classifiers based on the voting optimization data. One or more network nodes that host a voting classifier in the set of one or more selected voting classifiers is then notified of the selection.
A computer network is a geographically distributed collection of nodes interconnected by communication links and segments for transporting data between end nodes, such as personal computers and workstations, or other devices, such as sensors, etc. Many types of networks are available, ranging from local area networks (LANs) to wide area networks (WANs). LANs typically connect the nodes over dedicated private communications links located in the same general physical location, such as a building or campus. WANs, on the other hand, typically connect geographically dispersed nodes over long-distance communications links, such as common carrier telephone lines, optical lightpaths, synchronous optical networks (SONET), synchronous digital hierarchy (SDH) links, or Powerline Communications (PLC) such as IEEE 61334, IEEE P1901.2, and others. In addition, a Mobile Ad-Hoc Network (MANET) is a kind of wireless ad-hoc network, which is generally considered a self-configuring network of mobile routers (and associated hosts) connected by wireless links, the union of which forms an arbitrary topology.
Smart object networks, such as sensor networks, in particular, are a specific type of network having spatially distributed autonomous devices such as sensors, actuators, etc., that cooperatively monitor physical or environmental conditions at different locations, such as, e.g., energy/power consumption, resource consumption (e.g., water/gas/etc. for advanced metering infrastructure or “AMI” applications) temperature, pressure, vibration, sound, radiation, motion, pollutants, etc. Other types of smart objects include actuators, e.g., responsible for turning on/off an engine or perform any other actions. Sensor networks, a type of smart object network, are typically shared-media networks, such as wireless or PLC networks. That is, in addition to one or more sensors, each sensor device (node) in a sensor network may generally be equipped with a radio transceiver or other communication port such as PLC, a microcontroller, and an energy source, such as a battery. Often, smart object networks are considered field area networks (FANs), neighborhood area networks (NANs), personal area networks (PANs), etc. Generally, size and cost constraints on smart object nodes (e.g., sensors) result in corresponding constraints on resources such as energy, memory, computational speed and bandwidth.
Data packets 140 (e.g., traffic and/or messages) may be exchanged among the nodes/devices of the computer network 100 using predefined network communication protocols such as certain known wired protocols, wireless protocols (e.g., IEEE Std. 802.15.4, WiFi, Bluetooth®, etc.), PLC protocols, or other shared-media protocols where appropriate. In this context, a protocol consists of a set of rules defining how the nodes interact with each other.
The network interface(s) 210 contain the mechanical, electrical, and signaling circuitry for communicating data over links 105 coupled to the network 100. The network interfaces may be configured to transmit and/or receive data using a variety of different communication protocols. Note, further, that the nodes may have two different types of network connections 210, e.g., wireless and wired/physical connections, and that the view herein is merely for illustration. Also, while the network interface 210 is shown separately from power supply 260, for PLC (where the PLC signal may be coupled to the power line feeding into the power supply) the network interface 210 may communicate through the power supply 260, or may be an integral component of the power supply.
The memory 240 comprises a plurality of storage locations that are addressable by the processor 220 and the network interfaces 210 for storing software programs and data structures associated with the embodiments described herein. Note that certain devices may have limited memory or no memory (e.g., no memory for storage other than for programs/processes operating on the device and associated caches). The processor 220 may comprise hardware elements or hardware logic adapted to execute the software programs and manipulate the data structures 245. An operating system 242, portions of which are typically resident in memory 240 and executed by the processor, functionally organizes the device by, inter alia, invoking operations in support of software processes and/or services executing on the device. These software processes and/or services may comprise a routing process/services 244 and an illustrative “learning machine” process 248, which may be configured depending upon the particular node/device within the network 100 with functionality ranging from intelligent learning machine processes to merely communicating with intelligent learning machines, as described herein. Note also that while the learning machine process 248 is shown in centralized memory 240, alternative embodiments provide for the process to be specifically operated within the network interfaces 210.
It will be apparent to those skilled in the art that other processor and memory types, including various computer-readable media, may be used to store and execute program instructions pertaining to the techniques described herein. Also, while the description illustrates various processes, it is expressly contemplated that various processes may be embodied as modules configured to operate in accordance with the techniques herein (e.g., according to the functionality of a similar process). Further, while the processes have been shown separately, those skilled in the art will appreciate that processes may be routines or modules within other processes.
Routing process (services) 244 contains computer executable instructions executed by the processor 220 to perform functions provided by one or more routing protocols, such as proactive or reactive routing protocols as will be understood by those skilled in the art. These functions may, on capable devices, be configured to manage a routing/forwarding table (a data structure 245) containing, e.g., data used to make routing/forwarding decisions. In particular, in proactive routing, connectivity is discovered and known prior to computing routes to any destination in the network, e.g., link state routing such as Open Shortest Path First (OSPF), or Intermediate-System-to-Intermediate-System (ISIS), or Optimized Link State Routing (OLSR). Reactive routing, on the other hand, discovers neighbors (i.e., does not have an a priori knowledge of network topology), and in response to a needed route to a destination, sends a route request into the network to determine which neighboring node may be used to reach the desired destination. Example reactive routing protocols may comprise Ad-hoc On-demand Distance Vector (AODV), Dynamic Source Routing (DSR), DYnamic MANET On-demand Routing (DYMO), etc. Notably, on devices not capable or configured to store routing entries, routing process 244 may consist solely of providing mechanisms necessary for source routing techniques. That is, for source routing, other devices in the network can tell the less capable devices exactly where to send the packets, and the less capable devices simply forward the packets as directed.
Learning machine process 248 contains computer executable instructions executed by the processor 220 to perform various functions, such as attack detection and reporting. In general, machine learning is concerned with the design and the development of techniques that take as input empirical data (such as network statistics and performance indicators), and recognize complex patterns in these data. One very common pattern among machine learning techniques is the use of an underlying model M, whose parameters are optimized for minimizing the cost function associated to M, given the input data. For instance, in the context of classification, the model M may be a straight line that separates the data into two classes such that M=a*x+b*y+c and the cost function would be the number of misclassified points. The learning process then operates by adjusting the parameters a,b,c such that the number of misclassified points is minimal. After this optimization phase (or learning phase), the model M can be used very easily to classify new data points. Often, M is a statistical model, and the cost function is inversely proportional to the likelihood of M, given the input data.
As also noted above, learning machines (LMs) are computational entities that rely on one or more machine learning processes for performing a task for which they haven't been explicitly programmed to perform. In particular, LMs are capable of adjusting their behavior to their environment. In the context of LLNs, and more generally in the context of the IoT (or Internet of Everything, IoE), this ability will be very important, as the network will face changing conditions and requirements, and the network will become too large for efficiently management by a network operator.
Artificial Neural Networks (ANNs) are a type of machine learning technique whose underlying mathematical models that were developed inspired by the hypothesis that mental activity consists primarily of electrochemical activity between interconnected neurons. ANNs are sets of computational units (neurons) connected by directed weighted links. By combining the operations performed by neurons and the weights applied by the links, ANNs are able to perform highly non-linear operations to input data. The interesting aspect of ANNs, though, is not that they can produce highly non-linear outputs of the input, but that they can learn to reproduce a predefined behavior through a training process. Accordingly, an ANN may be trained to identify deviations in the behavior of a network that could indicate the presence of a network attack (e.g., a change in packet losses, link delays, number of requests, etc.).
Low power and Lossy Networks (LLNs), e.g., certain sensor networks, may be used in a myriad of applications such as for “Smart Grid” and “Smart Cities.” A number of challenges in LLNs have been presented, such as:
1) Links are generally lossy, such that a Packet Delivery Rate/Ratio (PDR) can dramatically vary due to various sources of interferences, e.g., considerably affecting the bit error rate (BER);
2) Links are generally low bandwidth, such that control plane traffic may generally be bounded and negligible compared to the low rate data traffic;
3) There are a number of use cases that require specifying a set of link and node metrics, some of them being dynamic, thus requiring specific smoothing functions to avoid routing instability, considerably draining bandwidth and energy;
4) Constraint-routing may be required by some applications, e.g., to establish routing paths that will avoid non-encrypted links, nodes running low on energy, etc.;
5) Scale of the networks may become very large, e.g., on the order of several thousands to millions of nodes; and
6) Nodes may be constrained with a low memory, a reduced processing capability, a low power supply (e.g., battery).
In other words, LLNs are a class of network in which both the routers and their interconnect are constrained: LLN routers typically operate with constraints, e.g., processing power, memory, and/or energy (battery), and their interconnects are characterized by, illustratively, high loss rates, low data rates, and/or instability. LLNs are comprised of anything from a few dozen and up to thousands or even millions of LLN routers, and support point-to-point traffic (between devices inside the LLN), point-to-multipoint traffic (from a central control point to a subset of devices inside the LLN) and multipoint-to-point traffic (from devices inside the LLN towards a central control point).
An example implementation of LLNs is an “Internet of Things” network. Loosely, the term “Internet of Things” or “IoT” may be used by those in the art to refer to uniquely identifiable objects (things) and their virtual representations in a network-based architecture. In particular, the next frontier in the evolution of the Internet is the ability to connect more than just computers and communications devices, but rather the ability to connect “objects” in general, such as lights, appliances, vehicles, HVAC (heating, ventilating, and air-conditioning), windows and window shades and blinds, doors, locks, etc. The “Internet of Things” thus generally refers to the interconnection of objects (e.g., smart objects), such as sensors and actuators, over a computer network (e.g., IP), which may be the Public Internet or a private network. Such devices have been used in the industry for decades, usually in the form of non-IP or proprietary protocols that are connected to IP networks by way of protocol translation gateways. With the emergence of a myriad of applications, such as the smart grid, smart cities, and building and industrial automation, and cars (e.g., that can interconnect millions of objects for sensing things like power quality, tire pressure, and temperature and that can actuate engines and lights), it has been of the utmost importance to extend the IP protocol suite for these networks.
An example protocol specified in an Internet Engineering Task Force (IETF) Proposed Standard, Request for Comment (RFC) 6550, entitled “RPL: IPv6 Routing Protocol for Low Power and Lossy Networks” by Winter, et al. (March 2012), provides a mechanism that supports multipoint-to-point (MP2P) traffic from devices inside the LLN towards a central control point (e.g., LLN Border Routers (LBRs) or “root nodes/devices” generally), as well as point-to-multipoint (P2MP) traffic from the central control point to the devices inside the LLN (and also point-to-point, or “P2P” traffic). RPL (pronounced “ripple”) may generally be described as a distance vector routing protocol that builds a Directed Acyclic Graph (DAG) for use in routing traffic/packets 140, in addition to defining a set of features to bound the control traffic, support repair, etc. Notably, as may be appreciated by those skilled in the art, RPL also supports the concept of Multi-Topology-Routing (MTR), whereby multiple DAGs can be built to carry traffic according to individual requirements.
A DAG is a directed graph having the property that all edges (and/or vertices) are oriented in such a way that no cycles (loops) are supposed to exist. All edges are contained in paths oriented toward and terminating at one or more root nodes (e.g., “clusterheads or “sinks”), often to interconnect the devices of the DAG with a larger infrastructure, such as the Internet, a wide area network, or other domain. In addition, a Destination Oriented DAG (DODAG) is a DAG rooted at a single destination, i.e., at a single DAG root with no outgoing edges. A “parent” of a particular node within a DAG is an immediate successor of the particular node on a path towards the DAG root, such that the parent has a lower “rank” than the particular node itself, where the rank of a node identifies the node's position with respect to a DAG root (e.g., the farther away a node is from a root, the higher is the rank of that node). Further, in certain embodiments, a sibling of a node within a DAG may be defined as any neighboring node which is located at the same rank within a DAG. Note that siblings do not necessarily share a common parent, and routes between siblings are generally not part of a DAG since there is no forward progress (their rank is the same). Note also that a tree is a kind of DAG, where each device/node in the DAG generally has one parent or one preferred parent.
DAGs may generally be built (e.g., by a DAG process) based on an Objective Function (OF). The role of the Objective Function is generally to specify rules on how to build the DAG (e.g. number of parents, backup parents, etc.).
In addition, one or more metrics/constraints may be advertised by the routing protocol to optimize the DAG against. Also, the routing protocol allows for including an optional set of constraints to compute a constrained path, such as if a link or a node does not satisfy a required constraint, it is “pruned” from the candidate list when computing the best path. (Alternatively, the constraints and metrics may be separated from the OF.) Additionally, the routing protocol may include a “goal” that defines a host or set of hosts, such as a host serving as a data collection point, or a gateway providing connectivity to an external infrastructure, where a DAG's primary objective is to have the devices within the DAG be able to reach the goal. In the case where a node is unable to comply with an objective function or does not understand or support the advertised metric, it may be configured to join a DAG as a leaf node. As used herein, the various metrics, constraints, policies, etc., are considered “DAG parameters.”
Illustratively, example metrics used to select paths (e.g., preferred parents) may comprise cost, delay, latency, bandwidth, expected transmission count (ETX), etc., while example constraints that may be placed on the route selection may comprise various reliability thresholds, restrictions on battery operation, multipath diversity, bandwidth requirements, transmission types (e.g., wired, wireless, etc.). The OF may provide rules defining the load balancing requirements, such as a number of selected parents (e.g., single parent trees or multi-parent DAGs). Notably, an example for how routing metrics and constraints may be obtained may be found in an IETF RFC, entitled “Routing Metrics used for Path Calculation in Low Power and Lossy Networks” <RFC 6551> by Vasseur, et al. (March 2012 version). Further, an example OF (e.g., a default OF) may be found in an IETF RFC, entitled “RPL Objective Function 0” <RFC 6552> by Thubert (March 2012 version) and “The Minimum Rank Objective Function with Hysteresis” <RFC 6719> by O. Gnawali et al. (September 2012 version).
Building a DAG may utilize a discovery mechanism to build a logical representation of the network, and route dissemination to establish state within the network so that routers know how to forward packets toward their ultimate destination. Note that a “router” refers to a device that can forward as well as generate traffic, while a “host” refers to a device that can generate but does not forward traffic. Also, a “leaf” may be used to generally describe a non-router that is connected to a DAG by one or more routers, but cannot itself forward traffic received on the DAG to another router on the DAG. Control messages may be transmitted among the devices within the network for discovery and route dissemination when building a DAG.
According to the illustrative RPL protocol, a DODAG Information Object (DIO) is a type of DAG discovery message that carries information that allows a node to discover a RPL Instance, learn its configuration parameters, select a DODAG parent set, and maintain the upward routing topology. In addition, a Destination Advertisement Object (DAO) is a type of DAG discovery reply message that conveys destination information upwards along the DODAG so that a DODAG root (and other intermediate nodes) can provision downward routes. A DAO message includes prefix information to identify destinations, a capability to record routes in support of source routing, and information to determine the freshness of a particular advertisement. Notably, “upward” or “up” paths are routes that lead in the direction from leaf nodes towards DAG roots, e.g., following the orientation of the edges within the DAG. Conversely, “downward” or “down” paths are routes that lead in the direction from DAG roots towards leaf nodes, e.g., generally going in the opposite direction to the upward messages within the DAG.
Generally, a DAG discovery request (e.g., DIO) message is transmitted from the root device(s) of the DAG downward toward the leaves, informing each successive receiving device how to reach the root device (that is, from where the request is received is generally the direction of the root). Accordingly, a DAG is created in the upward direction toward the root device. The DAG discovery reply (e.g., DAO) may then be returned from the leaves to the root device(s) (unless unnecessary, such as for UP flows only), informing each successive receiving device in the other direction how to reach the leaves for downward routes. Nodes that are capable of maintaining routing state may aggregate routes from DAO messages that they receive before transmitting a DAO message. Nodes that are not capable of maintaining routing state, however, may attach a next-hop parent address. The DAO message is then sent directly to the DODAG root that can in turn build the topology and locally compute downward routes to all nodes in the DODAG. Such nodes are then reachable using source routing techniques over regions of the DAG that are incapable of storing downward routing state. In addition, RPL also specifies a message called the DIS (DODAG Information Solicitation) message that is sent under specific circumstances so as to discover DAG neighbors and join a DAG or restore connectivity.
As noted above, LLNs are typically limited in terms of available resources and tend to be more dynamic than other forms of networks, leading to a number of challenges when attempting to detect DoS and other forms of network attacks. In particular, the limited computing resources available to a given network node may prevent the node from hosting a full-fledged learning machine process. In some cases, the node may simply export observation data to a learning machine hosted by a device with greater resources (e.g., a FAR). However, doing so also increases traffic overhead in the network, which may impact performance in an LLN.
According to various embodiments, lightweight learning machine classifiers may be distributed to network nodes for purposes of attack detection. In general, a classifier refers to a machine learning process that is operable to associate a label from among a set of labels with to an input set of data. For example, a classifier may apply a label (e.g., “Attack” or “No Attack”) to a given set of network metrics (e.g., traffic rate, etc.). The distributed classifiers may be considered “lightweight” in that they may have lower computational requirements than a full-fledged classifier, at the tradeoff of lower performance. To improve attack detection, a central computing device (e.g., a FAR, NMS, etc.) that has greater resources may execute a more computationally intensive classifier in comparison to the distributed lightweight classifier. In cases in which a distributed classifier detects an attack, it may provide data to the central device to validate the results and/or to initiate countermeasures. However, since the performance of a distributed classifier may be relatively low, this also means that there may be a greater amount of false positives reported to the central classifier.
Referring now to
To reduce the number of false positives, a voting mechanism may be implemented within network 100 to validate a detected attack before the supervisory device is notified. For example, as shown in
Voting Strategy Optimization Using Distributed Classifiers
The techniques herein provide mechanisms for computing an optimum voting strategy for a given classification problem between classifiers distributed across a network. In some aspects, network nodes/devices hosting classifiers that are potentially of interest are requested to apply their classifiers on a known validation set (e.g., a set of validation data containing known ground-truths). The classification results on this validation set may then be collected and used by an optimization process for computing an optimum voting strategy (e.g., which nodes/classifiers are to participate in a vote and the minimum value for the agreement between these classifiers). For example, the optimal voting strategy may be determined by minimizing or maximizing an objective function that is subject to one or more constraints. In further aspects, once the optimum voting strategy has been computed, the involved classifiers may be uploaded to the device that initiated the voting optimization process, allowing for local voting. The amount of information exchanged in the voting process after optimization may be reduced, making this approach well suited for applications such as the IoT.
Specifically, according to one or more embodiments of the disclosure as described in detail below, voting optimization requests that identify a validation data set are sent to a plurality of network nodes. Voting optimization data is received from the plurality of network nodes that was generated by executing classifiers using the validation data set. A set of one or more voting classifiers is then selected from among the classifiers based on the voting optimization data. One or more network nodes that host a voting classifier in the set of one or more selected voting classifiers is then notified of the selection.
Illustratively, the techniques described herein may be performed by hardware, software, and/or firmware, such as in accordance with the learning machine process 248, which may contain computer executable instructions executed by the processor 220 (or independent processor of interfaces 210) to perform functions relating to the techniques described herein, e.g., in conjunction with routing process 244. For example, the techniques herein may be treated as extensions to conventional protocols, such as the various PLC protocols or wireless communication protocols, and as such, may be processed by similar components understood in the art that execute those protocols, accordingly.
The following terms are introduced to aid in the understanding of the techniques herein:
Operationally, the techniques disclosed herein may involve performing any or all of the following functions. First, an LCE may send to other LCEs the characteristics of the classification problem and the characteristics of the validation set to which the LCEs should apply their classifier(s). Next, all the LCEs with a classifier and a validation set that satisfy the request (e.g., the classification and validation characteristics) may apply their classifier(s) to the validation set and send the results back to the requesting LCE. With the collected results, the requesting LCE may then compute the optimum voting strategy based on optimality criteria, such as how many correct classifications were performed by a pool of classifiers. The LCEs hosting classifiers selected to participate in the voting according to the computed optimal voting strategy are then contacted and may, in some embodiments, be requested to send their classifier to the LCE that launched the optimization. After reception of all the classifiers, the LCE that requested the voting optimization may be able to locally apply the optimum voting strategy computed, thereby reducing the amount of traffic used to conduct a vote. Alternatively, the remote classifiers may still be used to participate in a vote.
—Initiating a Voting Optimization—
Let LCE(i) be a LCE facing a classification problem using its local classifier C(i,j). To improve the results that LCE(i) obtains with C(i,j) applied alone, LCE(i) may request a voting optimization by sending requests to other LCEs. In general, the optimization requests may include the characteristics of the classification problem and the characteristics of the validation set that may be used to determine the optimum set of voting classifiers.
In one embodiment, a voting optimization request may be sent as an IPv4 or IPv6 unicast message to the NMS. For example, as shown in
In another embodiment, a voting optimization request may be sent to a known multicast group containing all of the available LCEs in the network. For example, as shown in
Voting optimization request 702 may include any or all of the following type-length-values (TLVs):
Upon reception of a voting optimization request, each receiving LCE(k) may check whether it has a classifier C(k,l) that satisfies the requested conditions (e.g., whether L(k,l)=L(i,j)). The receiving LCE(k) may also determine whether or not it has access to the validation set specified in the optimization request. In one embodiment, the access can be local to a memory of LCE(k). In another embodiment, LCE(k) may contact an NMS or a Network Controller and request access to the specified validation set. If any of these two conditions is not satisfied, LCE(k) may respond to the requesting LCE with an optimization refused message (e.g., an IPv4 or IPv6 message). In another embodiment, lack of acknowledgement from LCE(k) may be considered a negative reply by the requesting LCE.
If an LCE has a classifier that satisfies the optimization request, as well as access to the requested validation set, the LCE may perform the requested classification. In other words, the LCE may apply its local classifier(s) C(k,l) to every sample of the validation set V (e.g., the classifiers that satisfy the optimization request). If the optimization request includes constraints for the validation set, the LCE may first apply the constraints to validation set before applying its local classifier(s). For example, as shown in
Once the LCEs complete their evaluation of the validation set using their local classifiers, the LCEs may send the results back to the requesting LCE as voting optimization data. For example, as shown in
In another embodiment, a given LCE may opt to send its one or more local classifiers to the requesting LCE in addition to, or in lieu of, the voting optimization data. For example, assume that a responding LCE has a classifier C(k,l) that satisfies the conditions in the optimization request, but does not have access to the validation request. In such a case, the responding LCE may opt to send classifier C(k,l) to the requesting LCE. In this case, the requesting LCE may locally apply classifier C(k,l) to the validation set, to obtain the voting optimization data. Note that in cases in which all voting is performed locally by the requesting LCE (e.g., using different classifiers), classifier C(k,l) will already be resident on the requesting LCE at this point. Thus, a further request for the classifier will not be needed.
—Voting Optimization—
After reception of the voting optimization data (e.g., the results generated by the distributed classifiers on the validation set), the requesting LCE determines an optimal voting strategy for its classification problem. In cases in which a responding LCE sends its local classifier instead of the classifier's results, the requesting LCE may obtain the corresponding optimization data by apply the received classifier to the validation set. For example, as shown in
In various embodiments, the optimum voting strategy may be computed by an optimization process that takes as input all of the collected results of the distributed classifiers (e.g., the optimization data) and gives as output the optimum set of classifiers (N) and/or the optimum value of the voting agreement (k). These optimum values may be optimized with respect to a predefined optimality criterion. Such a criterion may be, for example, to maximize the number of well-classified samples, to maximize the number of well-classified samples of a particular class, to minimize the classification performance variation with respect to a particular parameter, to minimize a predefined weighted error measure, etc.
Voting optimization may be treated by a device as a discrete optimization problem that seeks to maximize or minimize an objective function subject to one or several constraints, and where all the variables involved take only natural values. For purposes of illustration, let f_V(k,N) be a function that counts the number of correctly classified samples in the validation set V considering a consensus of at least k classifiers in the set of classifiers N. For example, assume that k=2, N={C(i,j), C(k,l), C(m,n)}, and that the correct label for a particular sample in V is “attack.” In such a case, at least two of the classifiers in N may label the sample as “attack” for the consensus to be correct. If so, f_V(k,N) will increment the number of correctly classified samples. Otherwise, the consensus reached on the particular sample will be ignored by the function. In various embodiments, such a function may be treated as a discrete optimization problem that seeks to maximize f_V(k,N) subject to two conditions:
In other words, the number of eligible voters, as well as the vote count threshold for a consensus, may be reduced by the optimization problem, while still ensuring the number of correct voting results is maximized.
—Centralized Voting—
Once LCE(i) has computed N, the set of classifiers that may participate in the voting, LCE(i) may contact every LCE hosting one of the selected classifiers. For example, the requesting LCE may send a classifier request to each LCE that hosts one or more of the classifiers in N. For example, as shown in
In response to receiving a classifier request, an LCE may reply with a classifier grant message 804, if the requesting LCE(i) is granted access to the local classifier(s). In one embodiment, classifier grant message 804 includes the requested classifier(s), thereby allowing the requesting LCE to perform centralized voting going forward. In other words, the classifiers selected to vote may be executed locally by the requesting LCE after receipt of classifier grant messages 804. For instance, if a particular classifier is based on ANNs, the classifier grant message may contain the weights of the links between neurons, the activation function of the neurons, and any parameter required by these activation functions. Advantageously, performing all of the voting locally at the requesting LCE reduces network overhead in comparison to distributed voting, but at a tradeoff of requiring the use of additional resources by the LCE. If a consensus is reached, the LCE may then initiate corrective measures or generate an alert, such as alert 508 shown in
In some cases, an LCE may refuse to send its classifier to the requesting LCE and/or a distributed voting process may be used. For example, confidentiality issues, a policy defined in a policy engine, or other such factors may prevent the LCE from granting access to the requested classifier(s). In such a case, the refusing LCE may send a notification to the requesting LCE that identifies the classifier(s) for which access is not granted. On reception of a refusal, the requesting LCE may opt to exclude the restricted classifier(s) from the voting (e.g., by computing a new, optimized voting strategy). In some embodiments, a distributed voting mechanism may still be used even if the classifiers are not sent to the requesting LCE. For example, the requesting LCE may still employ a distributed voting mechanism, such as the process shown in
In some embodiments, a requesting LCE may be notified when any of the LCEs update a classifier selected as a voter. For example, the requesting LCE may set a parameter in classifier request 802 that requests notification of any relevant classifier updates. In such a case, the LCE receiving the request may add the requesting LCE to a subscription list. Each time a local classifier is updated, the LCE may then send to each of the subscribed LCEs the details of its new classifier (in this case, a proper reason code is included in the message). For example, as shown in
As shown in
Referring now to
At step 1015, voting optimization data is received from the other LCEs that received the optimization request, as described in greater detail above. Such optimization data may be based in part on a determination as to whether or not a responding LCE has a classifier that is compatible with the optimization request and whether or not the validation data indicated in the request is available. If both conditions are met, the responding LCE may use its eligible classifier(s) on the validation data. The results of the classifications may then be included in the voting optimization data that is returned to the requesting LCE. For example, the optimization data may identify the classifier(s), the results obtained by the classifier(s), and/or performance measurements regarding the classifiers or the results.
At step 1020, voters are selected, as described in greater detail above. In some embodiments, an optimal set of voters may be selected from among the classifiers indicated in the received optimization data. For example, an objective function may be optimized to select the set of voters/classifiers that were able to correctly classify the samples in the validation data the most number of times. In some embodiments, a vote count threshold may also be optimized as part of the objective function. In other words, the optimization may entail determining both the minimal set of voters that maximizes the number of correct results, as well as the optimal threshold to reach consensus that yielded the best performance.
At step 1025, a notification is sent to the nodes that host the classifiers selected as voters, as detailed above. For example, classifier requests may be sent to the other LCEs that requests access to the selected classifiers. In one embodiment, the other LCEs may respond by sending the selected classifiers back to the requesting LCE. The requesting LCE can then use these classifiers to perform a vote locally. In another embodiment, a distributed voting mechanism may still be employed in which case the classifier requests ask the other LCEs for access to the selected classifiers. Procedure 1025 then ends at step 1030.
At step 1115, the voting classifier may be received from another network node, as described in greater detail above. For example, in response to determining an optimal voting pool of classifiers that are located on other network nodes, a device may request the respective classifiers from the other nodes. In general, the received classifier may include any data needed to duplicate the remote classifier on the local device (e.g., parameters, input features, labels, etc.). For instance, if a particular classifier is based on ANNs, the received data may contain the weights of the links between neurons, the activation function of the neurons, and any parameter required by these activation functions.
At step 1120, as described in greater detail above, a local vote may be conducted using the received classifier. A local vote may be performed, for example, to validate a conclusion reached by a particular classifier. Notably, by performing the vote locally, local observation data may be used by the classifiers, thereby lowering the network usage by the voting process (e.g., in comparison to performing a distributed vote). If the vote validates the presence of an attack, the local device may then initiate further measures such as sending an alert to one or more other devices, making routing changes, etc. Procedure 1100 then ends at a step 1125.
It should be noted that while certain steps within procedures 1000-1100 may be optional as described above, the steps shown in
The techniques described herein, therefore, provide for the computation of an optimized voting strategy for a particular classification problem, such as classifying network traffic as being either normal or indicative of an attack. In effect, the voting classifiers act as a meta-classifier that may demonstrate improved performance over that of a single classifier. In addition, a vote may be conducted using different types of classifiers (e.g., ANNs, SVMs, naïve Bayesian, etc.), that have the same output labels (e.g., normal vs. attack) and may or may not have the same input features. Such a vote may be performed locally, thereby reducing network usage by a particular node, or may be performed in a distributed manner, thereby reducing the resource requirements of the node.
While there have been shown and described illustrative embodiments that provide for validating the detection of a network attack, it is to be understood that various other adaptations and modifications may be made within the spirit and scope of the embodiments herein. For example, while the techniques herein are described primarily with respect to attack-detection classifiers, the techniques herein may also be used to vote on different classification labels that are not related to attack detection (e.g., labels that relate to other network conditions). In addition, while the techniques herein are described primarily in the context of an LLN, the techniques herein may be applied more generally to any form of computer network, such as an enterprise network.
The foregoing description has been directed to specific embodiments. It will be apparent, however, that other variations and modifications may be made to the described embodiments, with the attainment of some or all of their advantages. For instance, it is expressly contemplated that the components and/or elements described herein can be implemented as software being stored on a tangible (non-transitory) computer-readable medium (e.g., disks/CDs/RAM/EEPROM/etc.) having program instructions executing on a computer, hardware, firmware, or a combination thereof. Accordingly this description is to be taken only by way of example and not to otherwise limit the scope of the embodiments herein. Therefore, it is the object of the appended claims to cover all such variations and modifications as come within the true spirit and scope of the embodiments herein.