The present invention relates to the field of detection of deviations in mobile application's network behavior. More particularly, the present invention relates to detection of malware activity in mobile networks.
Publications and other reference materials referred to herein, including reference cited therein, are incorporated herein by reference in their entirety and are numerically referenced in the following text and respectively grouped in the appended Bibliography which immediately precedes the claims.
Along with the significant growth in the popularity of smartphones and in the number of available mobile applications, the number of malware applications which harm users or compromise their private data is significantly increased. Furthermore, the significant growth of social networking and always-connected applications has caused a dramatically increasing influence on traffic and signaling loads on the mobile networks, potentially leading to network congestion incidents. Network overloads can be caused by either intended attacks or by benign, but unintentionally faultily designed, and thus “network unfriendly” applications. Both the malware activities and the “network unfriendly” applications regularly affect the network behavior patterns and can be detected by monitoring an application's network behavior. Thus, monitoring and analysis of network-active applications' traffic patterns is essential for developing effective solutions for the prevention of network overloads.
Traditionally Intrusion Detection Systems (IDS) are classified according to the protected system type as being either host-based (HIDS) or network-based (NIDS) [1]. A network-based IDS is located on a central or distributed dedicated server and monitors any number of hosts. Its performance is based on analysis of network related events, such as traffic volume, IP addresses, service ports, protocol usage, etc'. Traffic monitoring is usually accomplished at concentrating network units, such as switches, routers, and gateways. On the other hand, a host-based IDS resides on and monitors a single host machine. Its performance is based mainly on an analysis of events related to OS information, such as file system, process identifiers, system calls, etc' as disclosed in [8].
Many malware applications use network communication for their needs, such as sending a malicious payload or a command to a compromised device, or getting user's data from the device. Such types of behavior influence the regular network traffic patterns of the application and can be identified by learning the application's “normal” patterns and further monitoring network events.
Recently, with the dramatic increase in the number of malware applications targeting smartphones, various methods for intrusion detection on mobile devices have been proposed. Most of the IDSs for mobile devices have focused on host-based intrusion detection systems applying either anomaly or rule-based methods on the set of features that indicate the state of the device [17]. However, in most cases, the data interpretation processes are performed on remote servers motivated by limited computational resources of the mobile phone. Only a few of the proposed systems perform the learning or data analysis directly on the device [6, 10, 19] and even less have applied statistical or machine-learning techniques [10, 19], even though such techniques are very popular and have been successfully used in traditional anomaly detection systems [8, 19]. Most of the systems either send the observed data to the server for analysis [2, 4, 12, 14, 16, 22] or perform the learning process offline on the server and plant the learned models back to the devices for the detection process [15, 17, 18].
In a few earlier proposed systems the learning is performed on the mobile devices. For example, the system proposed by Shamili et al. [19] utilizes a distributed Support Vector Machine algorithm for malware detection on a network of mobile devices. The phone calls, SMSs, and data communication related features are used for detection. During the training phase support vectors (SV) are learned locally on each device and then sent to the server where SVs from all the client devices are aggregated. Lastly, the server distributes the whole set of SVs to all the clients and each of the clients updates his own SVs. Thus, although a part of the learning is performed on the device, the server and communication infrastructure, along with additional bandwidth load, are required.
Li et al. [10] presented an approach for behavior-based multi-level profiling IDS considering telephony calls, device usage, and Bluetooth scans. They proposed a host-based system which collects and monitors user behavior features on a mobile device. A Radial Basis Network technique was used for learning profiles and detecting intrusions. However, the system capabilities were, also, tested offline only using the MIT Reality dataset [5] and its feasibility on mobile devices was not tested or verified.
It is therefore a purpose of the present invention to provide a system for protecting mobile device users from harmful applications.
It is a further purpose of the present invention to provide a system for protecting cellular network infrastructure from targeted or benign overloads.
Further purposes and advantages of this invention will appear as the description proceeds.
In a first aspect the invention is a system for protecting mobile devices in cellular networks from unauthorized harmful applications and for protecting cellular network infrastructure from targeted or benign overloads comprising mobile cellular devices and a cellular network infrastructure, wherein some of the mobile devices comprise an application manager which is adapted to manage the aggregation and learning processes and a detection manager which is adapted to analyze network behavior and detect deviations, wherein the application manager and detection manager are adapted to monitor the applications running on a device, learn the patterns of mobile applications network behavior and detect meaningful deviations from the application's observed normal behavior, and where the cellular network infrastructure comprising a services module, a logic module and a database access unit adapted for aggregation and analysis of an application's network traffic patterns for numerous users.
In embodiments of the invention, the application manager comprises:
a Registration Unit, adapted to extract the list of all installed applications and device identifiers; a Features Extraction Manager, adapted to manage the extraction, aggregation and learning processes according to the defined time intervals, and report to the server the application's data according to the received schedule; a Models Manager, adapted to perform all the models related operations; a Configuration Manager, adapted to perform all the configuration related operations (i.e. load, get, update); a Logging unit, adapted to record the specified events in the log files; an Alerts Handler, adapted to present the alerts to user interface and report to the server; and a Communication Services unit, adapted to perform all the communication related operations with external systems (i.e. server).
In embodiments of the invention, the application manager comprises:
a Registration Unit, adapted to extract the list of all installed applications and device identifiers; a Features Extraction Manager, adapted to manage the extraction, aggregation and learning processes according to the defined time intervals, and report to the server the application's data according to the received schedule; a Models Manager, adapted to perform all the models related operations; a Configuration Manager, adapted to perform all the configuration related operations (i.e. load, get, update); a Logging unit, adapted to record the specified events in the log files; an Alerts Handler, adapted to present the alerts to user interface and report to the server; and a Communication Services unit, adapted to perform all the communication related operations with external systems (i.e. server).
In embodiments of the invention, the infrastructure comprises:
In embodiments of the invention, the services module comprises: a User's Registration unit, adapted to register the application on the new devices, and update the lists of installed applications according to the information received from the devices; a Features Distribution Manager, adapted to extract the features distribution process, and update the defined schedule; a Features Acquisition unit, adapted to acquire and store the features data from the devices; a Models Distributor, adapted to transfer collaborative models to the devices; and an Alerts Handler, adapted to send alerts to the devices, store it locally, and send relevant alerts to the system administrator.
In embodiments of the invention, the logic module comprises:
a Models Learner unit, adapted to induce the collaborative models representing applications traffic patterns for multiple users; a Models Manager, adapted to perform all the models related operations, and store and retrieve the models from the storage; and a Models-Change Detector, adapted to verify if a collaborative model has significant changes so that it needs to be updated on the devices.
In embodiments of the invention, the models related operations may comprise the following:
In a second aspect the invention is a method for protecting mobile devices in a cellular network from unauthorized harmful applications and for protecting cellular network infrastructure from targeted or benign overloads, wherein the mobile devices comprising an application manager and a detection manager adapted to learn the patterns of mobile applications network behavior and detect meaningful deviations from the application's normal behavior, and the cellular network infrastructure comprising a services module, a logic module and a database access unit adapted for aggregation and analysis of an application's network traffic patterns for numerous users, comprising the following steps:
In embodiments of the invention, if there is no “Anomaly” detected the process continues as usual.
In embodiments of the invention, if an “Anomaly” has been detected, a Models Manager unit (on the device) executes the relevant models related operations; initiates connection to the server in order to receive a collaborative model of the application in question.
In embodiments of the invention, the models related operations may comprise the following:
In embodiments of the invention, the method further comprises:
The invention provides a system invented for learning the patterns of mobile applications network behavior and detection of meaningful deviations from the application's normal network behavior. The main goal of this system is to protect mobile device users and cellular infrastructure from unauthorized malicious applications. The system is designed to (a) identify malicious attack or masquerading for an application which is already installed on a mobile device and (b) identify republishing of popular applications with injected malicious code. The detection is performed based on the application's network behavior patterns solely. For each application two types of models are learned. The first one represents the personal network usage pattern for a user. It is learned and resides on the user's device. The second one represents a common network usage pattern of numerous users. This model is learned and resides on the system server.
The performed data analysis reveals that applications have very specific network traffic patterns, and that certain application categories can be distinguishable from their traffic patterns. The system evaluation experiments were conducted with a wide range of different applications, their versions, and several self-developed and real malware applications. The results demonstrate that different levels of deviations from normal behavior can be detected accurately. Specifically, the deviation in up to 20-25 percent of observations might be due to variations in a user's behavior or an application's diverse functionality; deviations in various ranges from 0 up to almost 90 percent of instances might be observed due to an application's version update; lastly, the deviations in 60 and more percent of observations are regularly caused by injected malware. In addition, the conducted experiment described herein below proves the feasibility of the proposed implementation and analyses performance overhead on mobile devices.
The invention has therefore the following advantages:
All the above and other characteristics and advantages of the invention will be further understood through the following illustrative and non-limitative description of embodiments thereof, with reference to the appended drawings.
The invention serves two purposes. First, it allows protection of mobile device users from malware applications and second, allows for aggregation and analysis of an applications' network traffic patterns (to be used for development of solutions protecting cellular infrastructure from malicious and “network unfriendly” benign applications). Regarding the protection of users from malware applications, the system supports two main use cases. The first case relates to applications already installed on a device and the second, to newly downloaded and installed applications. In the first case, the network traffic pattern of an application can be changed due to: (a) the changes in the user's behavior or (b) an application update to a new benign version or (c) a malicious attack. In this case the system's purpose is to detect the deviation in the application's traffic pattern and correctly classify it to one of the three above mentioned reasons. In the second case, the system's purpose is to identify whether the new application is actually a modification of another application with some new (probably malicious) behavior.
For the above purposes the system follows the hybrid Intrusion Detection Systems (IDS) approach and is designed in the client-server architecture (the system may also work as a stand-alone client application, without the server side). The responsibility of the client-side software is to monitor the applications running on a device, learn their user-specific local models and detect any deviations from the observed “normal” behavior. The responsibility of the server-side software is the aggregation of data reported from mobile devices and the learning of collaborative models, which represent the common traffic patterns of numerous users for each application. The local models are used for detection of deviations in traffic patterns of installed applications; the collaborative models are used for verification of newly installed applications vs. the known traffic patterns.
A new application downloaded and installed on a device can be either a republished version of a popular application or an original application claiming to provide some new type of service. In both cases the new application can be benign or malicious. The malicious or cracked versions of the popular applications can be detected by comparison of their network patterns to those of the known applications. In the malicious version of a popular application the patterns are similar to those of the original application but with some level of deviation are expected to be detected.
The main processes of the system operate in the following way:
Initial User Registration
Management of Features Extraction Distribution
Local Learning
Features Reporting
Collaborative Learning
Anomaly Detection
New (Known) Application Handling
Unknown Application Handling
Module ‘2’ describes the application manager, comprising a Registration unit (21), responsible for the extraction of the list of all installed applications and device identifier; a Features Extraction Manager (22), responsible for managing the extraction, aggregation and learning processes according to the defined time intervals, and for reporting to the server the application's data according to the received schedule; an Alerts Handler (26), responsible for presenting the alerts to user interface and reporting to the server; a Models Manager (23), performing all the models related operations, such as deciding if there is enough data to start the learning process, getting the collaborative models from the server, storing them on the local storage, etc'; a Configuration Manager (24), responsible for loading and receiving (either from user or server) configuration parameters and updating the corresponding modules; a Logging (25), responsible for recording and logging the most important events in the log files; and a Communication Services (27), responsible for establishment and managing required communications with the server.
Module ‘3’ describes the detection manager, comprising a Features Extraction unit (31), that performs the measurements of the defined features at the defined time periods; a Features Aggregation unit (32), that is responsible for computing the specified aggregations over all the extracted measurements at the defined time periods. The instances of the aggregated data are used to induce machine-learning models representing an application's behavior and for further anomalies detection; a Local Learner unit (33), responsible for inducing the local models representing applications traffic patterns specific for the user; and an Anomaly Detector (34), responsible for online analysis of applications network behavior and detection of deviation from its normal pattern. A Features storage unit (41), responsible for storing the aggregated features of the monitored applications; and a Models storage unit (42), responsible for storing local and collaborative models of the monitored applications.
Module ‘7’ describes the Data-Base Access (DBA), a library with data access helpers (providing the basic functionality for data insertion, deletion, and update).
Module ‘8’ describes the Communication Services unit (81) responsible for establishment and managing required communications with clients; Logging unit (82) responsible for recording and logging the most important server events or errors in the log files; and Management & Monitoring unit (83) responsible for providing system's administrators and users data monitoring, analysis and management capabilities.
A UserApps storage unit (91), responsible for storing device identifier and the list of the installed applications for all registered users; a Features storage unit (92), responsible for storing the aggregated features of the monitored applications; and a Models storage unit (93), responsible for storing local and collaborative models of the monitored applications.
As mentioned above, the Features Extraction module (31) is responsible for extraction (i.e., measuring and capturing) of the defined list of features for each running application at each defined time period. For this purpose it uses the application-programming-interfaces (APIs) provided by the Android Software Development Kit (SDK). Below is a list of the extracted features in one embodiment of the invention:
The following features are additionally extracted in another embodiment of the invention:
The extraction time period is a configurable parameter. For the initial experiments with the system described herein below, it was set to 5 seconds, however it is subject to change according to the results of future evaluation experiments.
The purpose of the Features Aggregation module (32) is to provide a concise representation of the extracted application's traffic data. For this purpose, a list of various aggregation functions is defined. The instances of the aggregated data are used to induce machine-learning models representing an application's behavior and for further anomalies detection. To get a notion of the usefulness of the various features, an extended list of possible aggregated features is defined and evaluated. Below is a list of all the currently defined and aggregated features:
Similar to the extraction, the aggregation time period is a configurable parameter and it was set to 1 minute.
One of the main goals of the invention is to learn user specific network traffic patterns for each application and determine if meaningful changes occur in the application's network behavior. This task relates to the family of semi-supervised anomaly detection problems, which assumes that the training data has samples for “normal” data examples only. For the purpose of the present invention, the semi-supervised learning problem is converted into a set of supervised problems for which numerous well established and time efficient algorithms exist. For this purpose we follow the “cross-feature analysis” approach presented in [9], and then further analyzed by [13].
The main assumption underlying the “cross-feature analysis” approach is that in normal behavior patterns, strong correlations between features exist and can be used to detect deviations caused by abnormal activities. Thus, “cross-feature analysis” learns and explores the mutual correlations existing among different features. The basic idea of a cross-feature analysis method is to explore the correlation between one feature and all the other features. Formally, it tries to solve the classification problems Ci:{f1, . . . , fi−1, fi+1, . . . , fL}→{fi}, where {f1, f2, . . . , fL} in is the features vector and L is the total number of features. Such a classifier is learned for each feature i, where i=1, . . . L. Thus, an ensemble of learners for each one of the features represents the model through which each features vector will be tested for “normality”.
The anomaly detection module (34) is responsible for the online analysis of an application's network behavior and the detection of deviations from normal patterns. When a feature's vector representing a normal event is tested against Ci, there is a higher probability for the predicted value to match (for discrete features) or be very similar (for numeric features) to the observed value. However, in the case of a vector representing abnormal behavior, the probability of such a match or similarity is much lower. Therefore, by applying all the features models to a tested vector and combining their results, a decision about vector normality can be derived. The more different the predictions are from the true values of the corresponding features, the more likely that the observed vector comes from a different distribution than the training set (i.e., represents an anomaly event).
For each predictor Ci the probability of the corresponding feature value of a vector x to come from a normal event is computed. This probability, noted P(fi(x) is normal), is calculated as 1−distance(Ci(x), fi(x)), where Ci(x) is the predicted value and fi(x) is the actual observed value. The distance between two values for a single continuous feature is the difference in values divided by the mean of the observed values for that feature. If the difference is higher than mean value, the distance is assigned with a constant large value (such as 0.999). The distance for a discrete feature is the Hamming distance (i.e., 1 if the feature values are different and 0 if they are identical).
To get the total probability of a vector x to represent an normal event, a naive assumption about the sub-model's independence is made and then all the individual probabilities computed for each one of the feature values are multiplied. A threshold distinguishing between normal and anomalous vectors is learned during algorithm calibration on the data sets with labeled samples.
However, detection of abnormality in a single observed instance is not sufficient to determine whether that application's behavior has been meaningfully changed. Such sole anomalies can be caused by changes or noise in a user's behavior. In order to reduce the False Alarms rate and improve the effectiveness of the invention in general, a procedure which considers the consequent observations and derives a decision comprised of the individual predictions for each one of these observations is defined. For example, an alarm can be dispatched only when an anomaly was detected in a certain number of consequent instances (i.e., 3 consecutively observed instances were detected as anomalous) or when an anomaly was detected in a certain percent of instances during a specified time period (i.e., 3 or more anomalies during a 10 minute interval).
The invention is currently implemented for Android devices. However it can be also applied on other mobile operation platforms and on the network units as well, because its performance is based on network features alone.
The following examples are two-dimensional graphs of traffic patterns observed while analyzing data of several popular mobile applications with heavy network usage.
As can be seen from the graphs, each one of the analyzed applications has its own specific traffic pattern which is easily distinguishable from other applications (on each of the graphs, the axis value's range is different). Additionally, other features can be utilized for differentiation in less certain cases.
It can be seen from the graphs that different applications from the same functionality type have very similar traffic patterns among them, while the traffic patterns of different types of applications are different.
Based on the observations of the examples above, the following features of the invention are confirmed:
The calibration of the system serves several purposes: 1) selection of optimal features subset, 2) evaluation of several machine-learning algorithms as base learners, 3) determination of the minimal sufficient training set size, and 4) determination of the strategy for raising the “Anomaly” alarm in case one or more anomalous records are detected.
For evaluation of different classification algorithms and selection of features, the following standard measures are employed: True Positive Rate (TPR) measure (also known as Detection Rate), which determines the proportion of correctly detected changes from an application's normal behavior; False Positive Rate (FPR) measure (also known as False Alarm Rate), which determines the proportion of mistakenly detected changes in an actually normal application behavior; and Total Accuracy, which measures the proportion of a correctly classified application behavior as either anomalous or normal.
The purpose of testing the system was to evaluate the ability of the invention to distinguish between benign and malicious versions of the same application and between two benign yet different versions of the same application. Additionally, the low False Alarm rate on the data records of the same application version was verified.
For the calibration, a set of 16 datasets were extracted and prepared from the collected data (where the Features Extraction module was installed and ran on the personal Android devices of eight volunteer users, having from 2 weeks up to 3 months of data for each user). Each one of the 16 datasets consists of training and test records. In half of the datasets (i.e., in 8 datasets) both the training and test records are taken from the same version of a certain application. These datasets are used to verify a low detection rate on the records of the same application and determine the deviation level in traffic patterns that can be attributed to the application diversity and changes in a user's behavior. In the other 8 datasets, training and test records are taken from different versions of a certain application. These datasets are used to verify the higher detection rate than seen in the cases with the same application version. However, in some cases, the low detection rate for the different application versions is acceptable, as different application versions are not obligated to contain any network related updates.
For both, the calibration and testing of the system, the training size for all applications was limited to the maximum of 150 instances, and the test size to the maximum of 400 instances. On datasets with fewer available examples, the full training and test sets are utilized.
The initial set of defined aggregated features includes above 50 attributes. Extraction and aggregation of a large number of features on a mobile device is a very inefficient and resource wasting process. Additionally, learning classification models and detection with a large number of features is much more computationally expensive. Furthermore, the presence of redundant or irrelevant features may decrease the accuracy of the learning algorithm. Therefore, the purpose in the features selection is to identify a minimal set of the most useful features. There are several groups of features among the defined list of aggregated features for which extraction and calculation is performed together using the same amount of resources. Thus, reducing one or a few features from such a group, while at least one feature from such a group has to be calculated, will not reduce the extraction and calculation overhead significantly. The standard approaches for features selection, such as Filter and Wrapper, are not applicable in this case, as they cannot consider the above described constraints between the features. For this reason, twenty feature subsets of various sizes and containing various groups of features were manually defined. The threshold distinguishing between the normal and anomalous vectors is defined separately for each one of the features subset in the preliminary calibration, as it depends on the number and type of the involved features.
Considering the prevalence of numerical attributes among the defined aggregated features, and the resource consumption issue, the following classifiers were evaluated as candidates for the base-learner algorithm: Linear Regression, Decision Table, Support Vector Machine for Regression, Gaussian Processes for Regression, Isotonic Regression, and Decision/Regression tree (REPTree). The Weka open source library [23] was used for evaluation of these algorithms. All the defined feature subsets were tested with all the evaluated base learning algorithms on the calibration datasets described above.
As previously mentioned, sometimes abnormal instances can be caused by either changes in a user's behavior or by diversity in an application's functionality. In order to determine the acceptable rate of such abnormal instances in a normal application's behavior, the possible range between 5 and 25 percent with step of 5 was evaluated. Thus, the results of all the tested algorithms and feature subsets were evaluated for 5 different “anomaly acceptance” rates; 5, 10, 15, 20, and 25.
The results of the calibration reveal the two best combinations of the base learning algorithm and features subset. The two best base algorithms are the Decision Table and the REPTree. The two best features subsets, presented in Table 1, are very similar to each other; one of the subsets includes all the features from another plus two additional features.
As can be seen from the Table above, there are seven features included in both of the best subsets. Therefore, these features are the most useful for modeling application's network traffic.
As for the estimated algorithm accuracy performance, the Decision Table algorithm in conjunction with the features subset #1 and “anomaly acceptance” rate 20 percent results in TPR=0.8, FPR=0, and Total Accuracy=0.875 and the REPTree algorithm in conjunction with the features subset #2 and “anomaly acceptance” rate 25 percent demonstrates exactly the same accuracy values.
For a better insight into the detection rate observed in the calibration datasets, the results of the Decision Table algorithm in conjunction with the features subset #1 and the REPTree algorithm in conjunction with the features subset #2 are presented in Table 2 (errors are marked in underlined bold and italic font).
It can be seen that for most of the different application versions the detection rate is above the determined “anomaly acceptance” rate of 20-25 percent for both algorithms. At the same time, the detection rate on the test sets from the same application version is always below 20 percent. Thus, the detection strategy consisting of several steps can be defined as follows: 1) raise the “Anomaly Alarm” if at least 3 consecutive abnormal instances are detected, 2) raise the “Anomaly Alarm” if at least 3 abnormal instances are detected among the five consecutive observations, 3) raise the “Anomaly Alarm” if at least 3 abnormal instances are detected among the ten consecutive observations. According to this strategy, a system will raise an alert about any meaningful changes in an application's network patterns, including those caused by a version update. Further on the version update can be verified within the mobile OS and the Alert with the relevant information (including abnormal instances rate, whether a version update was detected and when) can be presented to the user.
An important question regarding the proposed detection system is how quickly the detection can be started (i.e., how many examples are needed for sufficient learning of the network traffic patterns)? To answer this question the two algorithms that gave the best results were evaluated using train sets of various sizes. This experiment was executed on all the calibration datasets, varying the training set size from 10 to 100 or the maximum of the available instances with step 10, and from 100 to 400 with steps of 25.
The results with both algorithms show that, in most cases, the training size of 30-50 examples is sufficient for learning a stable model which is able to determine the level of deviation between an application's traffic patterns correctly. However in several cases, for such diverse applications like Facebook and Gmail, a larger amount, such as 80-150 examples, is needed for learning a stable model. Considering the fact that in the experiments each data instance represents one minute of an application's network usage, the conclusion is that a relatively short time, varying from 30 minutes to 2.5 hours of network activity is required for the system to learn the patterns of a new application. Certain applications with rare network usage may actually require much longer time, while the required amount of network behavior data is aggregated.
To test the proposed system, a set of other 12 datasets, 6 with training and test records from the same application version and 6 with training and test records from different application versions, was used. Additionally, the system was tested with one self-written and five real malware applications. For the tests with the real malware, five infected applications and their benign versions were utilized. The infected applications and the corresponding versions of the benign application were retrieved from a repository collected by crawling the official and various alternative Android markets for over a year and a half. Two applications were injected with PJApps [20] Trojan; Fling and CrazyFish, two applications injected with Geinimi [21] Trojan; Squibble Lite and ShotGun, and one sample of DroidKungFu-B [11] malware found within the OpenSudoku game.
The PJApps Trojan, which was discovered in applications from unofficial Android marketplaces, creates a service that runs in the background, sends sensitive information containing the IMEI, Device ID, Line Number, Subscriber ID, and SIM serial number to a web server, and retrieves commands from a remote command and control server.
The Geinimi Trojan arrives on the device as part of repackaged version of legitimate applications. The applications repackaged with Geinimi Trojan have been found in a variety of locations, including unofficial marketplaces, file-share sites, and miscellaneous websites. When installed, the Trojan attempts to establish contact with a command and control server for instructions and once the contact is established, it transmits information from the device to the server and may be instructed to perform certain actions.
The DroidKungFu-B is a version of the DroidKungFu malware. The DroidKungFu-B version targets already rooted phones and requests for the root privilege. In either case (with or without the root privilege), the malware collects and steals the phone information (e.g., IMEI, phone model, etc.).
The detection rate of the Decision Table and REPTree algorithms in conjunction with the features subset #1 and #2 correspondingly, on the evaluated datasets are presented in Table 3 (detection errors are marked in underlined bold and italic font).
It can be seen that for all the malware applications, the high level deviations (60-100%) were detected. Furthermore, deviations at various levels were detected in most cases when the learned models were tested with instances from a different application version. The undetected versions of Facebook and WhatsApp applications can be explained by very few or no network-related changes in the considered application versions. Additionally, the detection rate for all the cases when the learned models were tested with instances from the same application version are below the defined “anomaly acceptance” rate of 20 percent for the Decision Table algorithm and of 25 percent for the REPTree algorithm. Thus, using the Decision Table algorithm's “anomaly acceptance” rate of 20 percent, the estimated method's accuracy on the test data is the following: TPR=0.82, FPR=0.0 and Total Accuracy=0.875. For the REPTree algorithm with the determined “anomaly acceptance” rate of 25 percent, the estimated accuracy on the test data is even higher: TPR=0.91, FPR=0.0, and Total Accuracy=0.94.
Considering the surprisingly high detection rate in the several real malware applications, in the self-written malware the 100% detection rate is not surprising, as the benign and malicious versions are significantly different in their network usage patterns. However, in the case of the real malware applications, the 100% detection rate is not obvious. In the applications infected with the Trojans, the main application's functionality is preserved and some new functionality is added. Thus, some part of the data related to the old functionality might be expected to remain unchanged. This is actually the case with the Fling application where online mobile advertisements are displayed while the application is in the phone's frontend in both versions. Thus, the records corresponding to the time when the game was actually played were less affected by the Trojan functionality and thus the observed detection rate is “only” 60%. Analysis of the data aggregated from the benign and malicious versions of the evaluated applications shows that the significant differences are caused by a background process that is running even when an application is not active and performs multiple connections (or connection attempts) with the server at constant time intervals. This behavior has a significant effect on such features, such as avg. sent\received bytes, number of sent\receive events, global outer\inner sent\receive intervals, and others. Most of the mentioned and significantly influenced features are contained in the utilized features subsets and this explains the high detection rate.
To evaluate the overhead caused by the learning and detection processes on mobile phone resources in terms of memory consumption, CPU load, and time needed for a model's induction and vector testing processes, experiments were performed on a Samsung Galaxy S GT-i9000 running Android OS version 2.2. One of the selected combinations, the REPTree algorithm in conjunction with the features subset #2, was used for the overhead evaluation experiments.
Online monitoring is performed for network-active applications only. Generally there are no more than 2-3 such applications running simultaneously on a device most of the time. Additionally, during the time periods of a user's normal activity, the number of such applications may reach no more than 10-15 network-active concurrent processes. Thus, for performance estimation, a scenario of 10 concurrently monitored applications is considered. The memory and CPU load were estimated for learning the 10 application models and further constant monitoring of their network traffic. For a better estimation of memory consumption, the results were averaged through 10 distinct experiments.
The memory consumption of the application changes in intervals from 7,035 KB±8 before the learning process has started to 7,272 KB±15 after storing the 10 learned models in memory (which is approximately 1.4% of the device's RAM). Storage of each additional model in memory consumes about 24 KB±0.7 on average. For comparison the memory consumption observed for several constantly running Android services and other popular applications is presented below: Android System—24,375 KB; Phone Dialer—8,307 KB; Antivirus—7,155 KB; TwLauncher—22,279 KB; Email—10,611 KB; and Gmail—9,427 KB. The detection process has no effect on the consumed memory.
The CPU consumption peaks occurred at the times of the actual model's learning and were in the interval of 13%±1.5. The model's learning operations occur very rarely, either when a new application is installed or when a model's update is needed (due to a new application version or changes in user's behavior). The CPU consumption observed during the process's idle time was in the interval of 0.7%±1.02. Time needed to learn a model (using 50 training examples) varies in intervals of 249 msec.±27.4.
The time needed for testing a single instance varies in intervals 3.6 msec.±2.5. Aggregated features vectors are tested once at the defined aggregation time interval (one minute for these experiments). The CPU consumed by testing 10 concurrent instances (one for each one of the assumed active applications) varies in intervals of 1.8%±0.8.
The results of this experiment depict the resources' overhead caused during the user's high activity time periods. During the, presumably much longer, time periods of the user's normal activity, an even lower overhead is expected.
Although embodiments of the invention have been described by way of illustration, it will be understood that the invention may be carried out with many variations, modifications, and adaptations, without exceeding the scope of the claims.
Number | Date | Country | Kind |
---|---|---|---|
222530 | Oct 2012 | IL | national |
Number | Name | Date | Kind |
---|---|---|---|
5991881 | Conklin et al. | Nov 1999 | A |
7603710 | Harvey et al. | Oct 2009 | B2 |
7607021 | Rayes et al. | Oct 2009 | B2 |
7823199 | Rathi et al. | Oct 2010 | B1 |
7854001 | Chen et al. | Dec 2010 | B1 |
8056132 | Chang et al. | Nov 2011 | B1 |
8695095 | Baliga et al. | Apr 2014 | B2 |
20060294588 | Lahann et al. | Dec 2006 | A1 |
20070097976 | Wood et al. | May 2007 | A1 |
20080016208 | Treinen | Jan 2008 | A1 |
20100054221 | Tang | Mar 2010 | A1 |
20110145920 | Mahaffey et al. | Jun 2011 | A1 |
20120151033 | Baliga et al. | Jun 2012 | A1 |
Entry |
---|
“Andromaly”: “a behavioral malware detection framework for android devices” Asaf Shabtai • Uri Kanonov. Yuvai ElovJci. Chanan Glezer. Yael Weiss; Revised: Oct. 27, 2010 ˜) Springer Science+Business Media, I,LC 2011. |
Amer, S., Hamilton, J. 2010. Intrusion Detection Systems (IDS) Taxonomy—A Short Review, Defense Cyber Security, 13 (2). |
Burguera, I., Zurutuza, U. and Nadjm-Tehrani, S. 2011. Crowdroid: Behavior-Based Malware Detection System for Android. In Proceedings of the 1st Workshop on Security and Privacy in Smartphones and Mobile Devices, CCCSSPSM'11. |
Chandola, V., Banerjee, A. and Kumar, V. 2009. Anomaly detection: A survey. ACM Comput. Surv. 41(3):1-58. |
Cheng, J., Wong, S.H., Yang, H., Lu, S. 2007. SmartSiren: virus detection and alert for smartphones. In Proceedings of the Fifth International Conference on Mobile Systems, Applications and Services. |
Eagle, N. and Pentland, A. S. 2006. Reality mining: Sensing complex social systems. Personal and Ubiquitous Computing, 10 (4):255-268. |
Enck, W, Gilbert, P., Chun, B.G., Cox, L.P., Jung, J., McDaniel, P. and Sheth, A.N. 2010. Taintdroid: an information-flow tracking system for realtime privacy monitoring on smartphones. In Proceedings of the 9th USENIX conference on Operating systems design and implementation (Berkeley, CA, USA, 2010). USENIX Association, 1-6. |
Felt, A.P, Finifter, M. , Chin, E., Hanna, S. and Wagner, D. 2011. A Survey of Mobile Malware in the Wild. In Proceedings of the 1st Workshop on Security and Privacy in Smartphones and Mobile Devices, CCS-SPSM'11. |
Garcia-Teodoro, P., Diaz-Verdejo, J., Macia-Fernandez, G., & Vazquez, E. 2009. Anomaly-based network intrusion detection: Techniques, systems and challenges. Comput. Secur. 28(1-2), 18-28. |
Huang, Y.-A., Fan, W., Lee, W. and Yu, P. S. 2003. Cross-feature analysis for detecting ad-hoc routing anomalies. In Proceedings of the 23rd International Conference on Distributed Computing Systems (Washington, DC, USA, 2003). ICDCS'03, IEEE Computer Society. |
Li, F. Clarke, N.L., Papadaki, M. and Dowland, P.S. 2010. Behaviour Profiling on Mobile Devices, International Conference on Emerging Security Technologies, (Canterbury, UK). 77-82. |
McAfee: Android/DroidKungFu.B http://home.mcafee.com/virusinfo/virusprofile.aspx?key=522721. |
Moreau, Y., Verrelst, H., Vandewalle, J. 1997. Detection of mobile phone fraud using supervised neural networks: a first prototype. In: Proceedings of the Seventh International Conference on Artificial Neural Networks. |
Noto K, Brodley C, Slonim D. 2010. Anomaly detection using an ensemble of feature models. In: Proceedings of the 10th IEEE international conference on data mining. |
Portokalidis, G., Homburg, P., Anagnostakis, K. and Bos, H. 2010. Paranoid android: versatile protection for smartphones. In Proceedings of the 26th Annual Computer Security Applications Conference, (New York, NY, USA, 2010). ACM, 347-356. |
Schmidt, A.D., Schmidt, H.G., Clausen, J., Yüksel, K. A., Kiraz, O., Camtepe, A. and Albayrak, S. 2008. Enhancing security of linux-based android devices. In Proceedings of 15th International Linux Kongress. |
Schmidt, A.D., Peters, F., Lamour, F., Scheel, C., Camtepe, S.A., Albayrak, S. 2009. Monitoring smartphones for anomaly detection. Mobile Netw. Appl. 14 (1), 92-106. |
Shabtai, A., Kanonov, U. and Elovici, Y. 2010. Intrusion detection for mobile devices using the knowledge-based, temporal abstraction method. J. Syst. Softw. 83:1524-1537. |
Shabtai, A., Kanonov, U., Elovici, Y., Glezer, C. and Weiss, Y. 2011. Andromaly: a behavioral malware detection framework for android devices. J. Intell. Inf. Syst. 1-30, 10.1007/s10844-010-0148-x. |
Shamili, A.S., Bauckhage, C. and Alpcan, T. 2010. Malware detection on mobile devices using distributed machine learning. In Proceedings of the 20th International Conference on Pattern Recognition (Washington, DC, USA, 2010). ICPR'10, IEEE Computer Society, 4348-4351. |
Symantec: Android.Pjapps http://www.symantec.com/security—response/writeup.jsp?docid=2011-022303-3344-99&tabid=2. |
Symantec: Android.Geinimi http://www.symantec.com/security—response/writeup.jsp?docid=2011-010111-5403-99&tabid=2. |
Qian, F., Wang, Z., Gerber, A., Mao, Z., Sen, S. and Spatscheck, O. 2011. Profiling resource usage for mobile applications: a cross-layer approach, In Proceedings of the 9th International Conference on Mobile Systems, Applications, and Services (Bethesda, Maryland, USA). |
Weka 3: Data Mining Software in Java http://www.cs.waikato.ac.nz/ml/weka/ (prior to Oct. 2012). |
Number | Date | Country | |
---|---|---|---|
20140113588 A1 | Apr 2014 | US |