Malware detection through distributed telemetry data analysis
A distributed malware detection system using local and federated neural networks analyzes telemetry data to efficiently identify and neutralize malicious processes, overcoming resource bottlenecks and disguising behaviors in centralized sandboxing.
Patent Information
- Application Number
- JP2023535644
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2020-10-13
- Filing Date
- 2021-09-27
- Publication Date
- 2024-11-06
- Estimated Expiration
- 2041-09-27
AI Technical Summary
Centralized sandboxing for malware detection requires excessive system resources and can be bottlenecked by the number of secure enclave executions, and advanced malware can disguise its behavior as legitimate processes, making it difficult to detect effectively.
A distributed malware detection system using locally trained neural networks and federated trained neural networks that analyze telemetry data to identify malicious processes, reducing resource requirements and adapting to new patterns through periodic retraining.
This approach significantly reduces the CPU, memory, and storage needed for malware detection, effectively identifying and neutralizing advanced malware without overloading central systems and adapting to evolving threats.
Smart Images

Figure 0007578828000003 
Figure 0007578828000004 
Figure 0007578828000005
Abstract
Description
[Technical field]
[0001] The present invention relates generally to the field of malware detection, and more particularly to detecting malicious processes by selected instances of an anti-malware system. [Background technology]
[0002] For many years now, cybersecurity attacks continue to be a top priority for government agencies and information technology (IT) organizations and companies. The role of Chief Information Security Officer (CISO) is a known one in any IT organization. Most of the security information and event monitoring (SIEM) solutions available in the market require a complex set of correlation rules monitored and tuned by highly trained personnel to identify and detect potential security threats that may lead to security incidents in at least some part of the organization. According to the General Data Security Regulation (GDPR), companies operating within the European Union may be subject to high fees if the company fails to protect customer and / or employee data. However, protecting corporate data is not only a topic for European-based companies and government agencies, but for all businesses that handle data around the world.
[0003] Typically, malware may be received via a network connection or a locally attached storage medium and may be activated by some event. For example, a program may be launched in a computing environment (i.e., as a process of an operating system). One method for controlling such potentially unacceptable processes is known as sandboxing, in which unknown processes are executed in a protected environment (i.e., a processing environment that does not infect the remaining processes, operating system, and data) and potentially removed without harming the underlying computing system. Often, malware sandboxing and unknown file / process execution are implemented as a centralized sandbox, typically in a virtual appliance, where unknown files and code are executed to understand the behavior of the unknown files and processes, to collect Indicators of Compromise (IoCs) associated with the unknown malware, and to further analyze the malware's behavior.
[0004] Such a centralized implementation may also be problematic because parallel secure enclave execution of unknown files may require more powerful CPUs, more memory, and potentially more storage, making the centralized malware sandboxing a bottleneck for testing unknown file execution. Also, a centralized implementation may limit the size and capabilities of the centralized sandbox, as the amount of system resources for parallel secure enclave execution may grow out of proportion.
[0005] In addition, advanced malware may also detect centralized sandbox execution and disguise malware behavior as a legitimate process. Therefore, in developing a centralized sandbox system for malware analysis and unknown files / processes, there is a high risk of bottlenecks and the amount of available system resources required in exchange for the number of secure enclave executions performed by the centralized sandbox system may be negligible.
[0006] Document US2019 / 0319977A1 discloses an apparatus, system and method for fingerprinting and classifying application behavior using telemetry. An example apparatus includes a trace processor that processes events in a processor trace to capture application execution behavior, and a fingerprint extractor that extracts a first fingerprint from the captured application execution behavior and performance monitoring information.
[0007] In addition, document US10,148,686B2 discloses a system, method and computer program for analyzing telemetry data from physical process sensors to detect quantities in a physical process. The telemetry analysis system is disclosed as a process-level anomaly detection system based on behavioral telemetry and domain-specific knowledge that protects cyber-physical system devices against zero-day exploits that cannot be detected via other system logs or network packet inspection.
[0008] A problem with using centralized sandboxing, however, is that it may require excessive resources of the host system. Thus, there is a need to overcome this bottleneck and provide a method to (i) identify malicious processes on a host that may not be detected by traditional malware detection systems, and (ii) that requires significantly fewer resources than centralized sandboxing approaches. Summary of the Invention
[0009] Aspects of the invention disclose methods, computer program products, and systems for detecting malicious processes by a selected instance of an anti-malware system. The method includes one or more processors inspecting a process for indicators of compromise for the process. The method further includes one or more processors determining a classification of the process based on results of the inspection. In response to determining that the classification of the process does not correspond to a known good process and a known malicious process, the method further includes one or more processors executing the process within a secure enclave. The method further includes one or more processors collecting telemetry data from the process executing within the secure enclave. The method further includes one or more processors passing the collected telemetry data to a locally trained neural network system. The training data for the locally trained neural network system includes telemetry data from a process executing on a host system underlying the locally trained neural network system. The method further includes one or more processors determining a result of a first loss function for the locally trained neural network system. The method further includes the one or more processors comparing a result to a result of the loss function at the end of training of the locally trained neural network system.
[0010] In a further aspect, the method includes one or more processors passing the collected telemetry data to a federated trained neural network system. The federated trained neural network system is adapted to receive the federated trained neural network model. The method further includes the one or more processors determining a result of a second loss function for the federated trained neural network system. The method further includes the one or more processors comparing the result to a result of the second loss function of the received federated trained neural network model of the federated trained neural network system. The method further includes the one or more processors aggregating the results of the first loss function and the second loss function. The method further includes the one or more processors determining whether the process is anomalous based on the aggregated result.
[0011] The proposed method for detecting malicious processes by a selected instance of an anti-malware system may provide multiple advantages, technical effects, contributions and / or improvements.
[0012] First, a single centralized approach to detecting malicious processes may be avoided. Thus, the system resources required to detect malicious processes may be significantly reduced when compared to conventional approaches. Thus, the amount of central processing unit (CPU) power, memory and storage required to implement a sensitive intrusion detection system on a particular host may be reduced.
[0013] The use of telemetry data in addition to detecting traditional indicators of compromise may be tailored to individual environments, assigning different weights to different components of the telemetry data. In addition, the telemetry data may have static and dynamic components, which may be rapidly expanded as new methods and computing environments for deriving process telemetry data become available. The underlying process of using an autoencoder to detect potentially malicious patterns in a process may remain unchanged.
[0014] The option of periodic retraining of the locally trained or federated trained neural network system using separately trained neural network models can allow rapid adaptation to newly available telemetry data patterns.
[0015] In addition, the combined knowledge of the selected instances of the malware detection system in the form of a federated trained neural network model using the integrated "experience" of each of the individual instances of the locally trained neural network system and associated neural network model can achieve two goals. First, the analysis of process parameters (i.e., telemetry data) of the multiple host systems can be performed in a distributed manner so that the central instance is not overloaded with too much data to process in a limited time. Second, telemetry data from different processing environments (i.e., processor types, available memory and storage workload amounts and workload types, etc.) can be considered for training the federated neural network model. Once deployed and periodically updated, the federated trained neural network model may be redistributed to the selected instances of the malware detection system for use in the federated trained neural network system.
[0016] Thus, this aspect of the invention makes the method and related system proposed herein applicable, relatively easy to implement, scalable, and flexible, since resources of different hosts can be used in combination with the central instance without overloading the central instance and without capturing analytical data from the local instances. However, on the other hand, the local instances do not completely depend on the central instance (i.e., the training environment for the federated neural network model), but have available local resources in the form of locally trained neural network systems and federated trained neural network systems to defend the corresponding computing environment against malicious processes. Thus, the "central cyber attack defense intelligence" and the local protection system can be combined in a powerful and unique way.
[0017] Additionally, aspects of the present invention are operable to detect centralized sandbox execution and to detect and remove advanced malware that can disguise its behavior as legitimate process behavior. Thus, this goal may be achieved using limited resources of selected instances of a federated backbone of a malware detection system.
[0018] In the following, additional embodiments of the inventive concepts are described that are applicable to methods and also to associated computer systems and computer program products.
[0019] In another aspect, the method further includes one or more processors collecting telemetry data of non-malicious processes running on the selected instance of the anti-malware system. The method includes one or more processors retraining the locally trained neural network system with the collected telemetry data to periodically build an updated local neural network model. Thus, the non-malicious processes are processes that run periodically on the host under the control of the operating system. The non-malicious processes may be operating system processes or application program processes. On a typical host system, hundreds or thousands of such processes may run in parallel every second. The collection of telemetry data may be supported by a system management tool. Alternatively, the data may be collected independently using a component of a computer security framework. The corresponding generated telemetry data of the non-malicious processes (i.e., normal processes) may be used at a predetermined time (e.g., periodically or based on a trigger event) to retrain the local neural network model of the local neural network system. Thus, changes in workload and telemetry data profiles that may change over time may be reflected in the behavior of the local neural network model.
[0020] Additionally, the method further includes the one or more processors receiving an updated federated neural network model for the federated trained neural network system. The federated neural network model is constructed using the locally trained neural network systems of the selected instances as input. Thus, the federated neural network model may be constructed with the locally trained neural network models of the selected instances as input. Thus, the collective, cooperative "knowledge" of multiple hosts and their associated malware detection methods and systems that may be subject to malware attacks of various profiles may be the basis for group-wide "immunity" against various types of cyber attacks.
[0021] Moreover, at least one of the locally trained neural network system and the federated trained neural network system is an autoencoder system. Autoencoders are particularly well equipped to support the proposed concept. Since the number of artificial neurons (or nodes) in the input layer of each neural network is equal to the number of artificial neurons in the output layer, aspects of the present invention can operate to apply and minimize the result of a loss function based on the difference between the input layer and the output layer. However, if the input values to the neural network are comparatively close to the training telemetry data used, a minimum value may be reached, or the result of the loss function may be within a range of a minimum value. If other telemetry data is outside the range of the training telemetry data that may be applied to the input layer of the respective neural network, the loss function may exhibit a value greater than a predetermined threshold, suggesting the presence of a potentially malicious process.
[0022] In another aspect, the method further includes, in response to determining that the classification of the process corresponds to an anomalous process, the one or more processors discarding the process. Accordingly, processes based on potential malware program code may not leave the secure enclave and potentially in fact other processes or manipulate data of the host computing system in an unallowable manner.
[0023] In another aspect, in response to determining that the process is a normal process, based on execution in the secure enclave, the method further includes the one or more processors moving the process outside the secure enclave. The method includes the one or more processors executing the process as a normal process. Thus, the process may be executed as a normal process controlled by the operating system, potentially providing a basis for generating additional telemetry data of the normal process.
[0024] In an additional aspect, the federated trained neural network system is trained using telemetry data from multiple hosts. For example, the method may operate to transmit locally trained neural network models (e.g., from each local host) (e.g., in the form of corresponding parameters and hyperparameters) to a centralized system, such as one operating on a cloud computing system.
[0025] Training the federated trained neural network system further includes one or more processors processing the representative set of telemetry data by each of the plurality of received locally trained neural network models and storing respective results, and one or more processors training the federated neural network model using telemetry data input / output pairs used and generated during processing of the telemetry data as input data for training the federated neural network model. The telemetry data input / output pairs used to train the federated neural network model are weighted depending on the geographical proximity between the geographical host location of the locally received neural network model and the geographical host location where the new federated neural network model is trained. The telemetry data input / output pairs used to train the federated neural network model are weighted depending on a predefined metric.
[0026] This allows local details of individual host systems and corresponding federated trained neural network models to be reflected. An embodiment of the present invention recognizes that cyber attacks of a certain profile are often targeted to a specific geographical area. The proposed concept allows such specificities to be better reflected in the malware detection system, which can increase susceptibility to certain cyber attacks and advantageously save system resources. Thus, individual federated trained neural network models on different hosts may be different in such an embodiment.
[0027] In an exemplary embodiment, each of the 10 instances may send a corresponding local neural network model to a central service of the malware detection system. The geographic location of each may be known. Then, for example, a process for training a federated neural network model of instance 3 would include: (i) evaluating distance information between instance 3 and all other 9 instances; (ii) processing each representative set of telemetry data by the received locally trained neural network model and storing each result (as described above); and (iii) repeating the use of input / output pairs inversely proportional to the distances determined under step (i).
[0028] In an additional exemplary aspect, when considering a pair from instance 5, which is geographically very close to instance 3, the pair may be repeated 10 times (a predetermined number of times) during training (i.e., biasing the network); instead, when considering a pair from instance 7, which is, for example, 10 times further away, the value pair will only appear once in the training data.
[0029] According to another embodiment of the method, the input / output pairs of telemetry data used to train the federated neural network model may be weighted depending on predefined metrics. The weighting may allow to reflect not only the geographic location and its details in relation to a potential cyber attack, but also other profiles of the local host. For example, use within a certain vertical industry, a specific department, a type of host computer (e.g., a central system corporate computing system versus an Internet of Things (IoT) system or a personal device). Furthermore, the weighting may allow improved protection against malware resulting in potentially malicious processes depending on more adjustable parameters for malware.
[0030] According to another embodiment of the method, aggregating the results of the loss functions may include constructing a weighted average of the loss functions, which may reflect that the values of the respective weighting coefficients may be adjusted according to the amount of data seen by the federated trained neural network model (e.g., a central autoencoder) versus the locally trained neural network model (e.g., a local autoencoder). Thus, a "more experienced" autoencoder may have "more influence."
[0031] In another aspect, the method of aggregating results of the first and second loss functions further includes the one or more processors constructing a weighted average of the first and second loss functions. In another aspect, the determined classification is selected from the group consisting of a known good process, a known bad process, and an unknown process.
[0032] Furthermore, embodiments may take the form of an associated computer program product accessible from a computer usable or computer readable medium that provides program code for use by or in connection with a computer or any instruction execution system. For purposes of this specification, a computer usable or computer readable medium may be any apparatus that can encompass means for storing, communicating, propagating or transporting a program for use by or in connection with an instruction execution system, apparatus, or device.
[0033] The above-defined and further aspects of the invention will be apparent from and will be elucidated with reference to the example embodiments described hereinafter, to which the invention is not limited. Preferred embodiments of the invention will now be described, by way of example only, with reference to the following drawings, in which: [Brief description of the drawings]
[0034] [Figure 1]FIG. 1 illustrates a block diagram of an embodiment of a method for detecting malicious processes by a selected instance of an anti-malware system, according to an embodiment of the present invention. [Diagram 2] FIG. 2 illustrates an example of telemetry measurement data for a non-abnormal process, according to an embodiment of the present invention. [Diagram 3] FIG. 3 illustrates an exemplary structure of an autoencoder neural network according to an embodiment of the present invention. [Figure 4] FIG. 4 shows a first part of a more detailed exemplary flowchart of the proposed concept according to an embodiment of the present invention. [Diagram 5] FIG. 5 shows a second part of a more detailed exemplary flowchart of the proposed concept according to an embodiment of the present invention. [Figure 6] FIG. 6 illustrates a block diagram of an embodiment of components of the concepts proposed herein and corresponding dependencies, according to an embodiment of the present invention. [Figure 7] FIG. 7 illustrates a block diagram of an embodiment showing a process for generating telemetry and metadata for a given process to be examined, according to an embodiment of the present invention. [Figure 8] FIG. 8 illustrates a block diagram of an embodiment showing a partial process for anomaly detection, including a user of a trained neural network module, in accordance with an embodiment of the present invention. [Figure 9] FIG. 9 shows a block diagram of an embodiment of the proposed malware detection system according to an embodiment of the present invention. [Figure 10] FIG. 10 shows a block diagram of an embodiment of a computing system including the proposed malware detection system according to FIG. 9, according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0035] Below, a detailed description of the drawings will be given. All indications in the figures are schematic. First, a block diagram of an embodiment of the inventive method for detecting a malicious process by a selected instance of an anti-malware system is given. After that, further embodiments will be described, as well as an embodiment of a malware detection system for detecting a malicious process by a selected instance of an anti-malware system.
[0036] It should be noted that embodiments of the present invention are described with reference to different subject matters. In particular, some embodiments are described with reference to method type claims, and other embodiments are described with reference to device type claims. However, a person skilled in the art will understand from the above and the following description that, unless otherwise specified, in addition to any combination of features belonging to one type of subject matter, any combination of features relating to different subject matters, in particular between features of a method type claim and a feature of an device type claim, is also considered to be disclosed in this document.
[0037] In the context of this specification, the following conventions, terms and / or expressions may be used.
[0038] The term "malicious process" refers to and may relate to activity associated with any computer code on a computing system that compromises the operational integrity of the computing system and the operating system and the normal application of the computer system. A malicious process may be based on data / computer code received on the host system (i.e., the computing system). Thus, an abnormal or non-regular process (i.e., a deliberate process executed under the control of the operating system) may be classified as a malicious process. A malicious process may modify data on the computing system, may observe or surreptitiously snoop on data on the computing system, and may alter any functionality of the computer system in an unintended manner.
[0039] The term "selected instance" may refer to a computing system or host that is part of the anti-malware system. The selected instance may also be referred to as the local host.
[0040] The term "anti-malware system" may refer to multiple selected instances working in cooperation with a central instance that supports the functionality of the local instances for better protection against cyber attacks.
[0041] The term "indicators of compromise" (IoC) may refer to a well-known concept in computer forensics and may represent artifacts observed on a network or within an operating system that may indicate a computer intrusion with a high degree of confidence. Types of indicators of compromise may include virus signatures, known suspicious Internet Protocol (IP) addresses, MD5 hash values of malware files, Uniform Resource Locators (URLs) of domain names of botnet command and control servers, etc. After identifying IoCs via incident response and / or computer forensics processes, embodiments of the present invention may utilize IoCs for early detection of future attacks using intrusion detection systems and / or antivirus software.
[0042] The term "known benevolent process" refers to a process in an operating system that has a high degree of trust and that does not harm the operating system in any way, as opposed to a malicious process. For example, a known benevolent process may be purposefully initiated by the operating system as part of the operating system or an associated application software program. A known benevolent process may also relate to program code received over a network from a secure source.
[0043] The term "known malicious process" may refer to a process in an operating system that has a high degree of confidence in compromising the operating system or the underlying computing system. In various aspects, IoCs may be used to identify a known malicious process.
[0044] The term "unknown process" may refer to a process in an operating system that is not classified as a malicious process or as a normal / secure process.
[0045] The term "secure enclave" may refer to a secure sandbox computing environment that can protect the normal operating environment of a computing system and / or its respective operating system and its applications.
[0046] The term "telemetry data" may refer to at least two classes of character arising data of an executing process: (a) static telemetry data and (b) dynamic telemetry data. Static telemetry data relates to data about an addressed process (i.e., program code defining a process that has not yet been executed). Static telemetry data may include at least (i) determined file hash values of executable files for file fingerprinting, (ii) string extraction for string analysis from executable files, (iii) layout extraction of executable files, and (iv) automated disassembly of code and code structures of executable files. In contrast, dynamic telemetry data may be data of an observed, executing process. Dynamic telemetry data may include at least (i) system monitoring data, (ii) registry activity data, (iii) file (i.e., file access) activity data, (iv) process activity data, and (v) network activity data. Thus, dynamic telemetry data may change over the life of the process and may indicate a characteristic profile.
[0047] This often involves the use of known techniques. For example, the following can be mentioned: String extraction is one of the first things to conveniently extract useful information such as IP addresses, domains, functions, data, or any other information that has not been removed by the malware developer. Hereby, a string is a sequence of characters, which may be a constant or a variable. A string is a type of data, which may be implemented as a byte (or word) data structure that stores a sequence of elements. Strings are inherent to computer programming, and are essentially valuable data used to create software and to gain clues about functionality in reverse engineering. Layout extraction focuses on checking the headers of executable program code to map information about the code, the type of application, the library functions required, and the space requirements of potential malware. File hashing, which is done initially to fingerprint malware executables, is typically done to extract the MD5 of the malware file, and is not suitable for extracting the details mentioned in this step. Automated disassembly of code and code structure may be performed by known or to be developed tools that can quickly disassemble code. A common tool used for this interactive disassembly for static malware analysis.
[0048] The term "locally trained neural network system" may refer to a neural network system that uses locally available training data (e.g., locally generated telemetry data of a normal process) to develop a neural network model. The neural network model and / or the neural network system may be constructed according to the rules of an autoencoder neural network. However, other types of neural networks may also be advantageously used.
[0049] The term "host system" may refer to a computing system, also referred to as a selected instance within a malware detection system.
[0050] The term "federated trained neural network system" may refer to a neural network system that includes an underlying neural network model that is trained on a central instance of a malware detection system. The federated trained neural network system may be operated as a cloud service and can provide support for malware detection and protection services for multiple local instances or host systems.
[0051] The term "first loss function" may refer to a function that returns a result value or a vector of results when input values to a neural network (e.g., an autoencoder) are compared with the respective output values. In particular, for an autoencoder, the output values may always be more or less different from the respective input values (e.g., because no autoencoder has an underlying idea and a perfect model). The first loss function may be associated with a trained local neural network system, while the second loss function may be associated with a federated trained neural network system, which may also be an autoencoder.
[0052] The term "updated federated neural network model" may refer to a new federated neural network model that is retrained with new training data. The updated federated neural network model may then be distributed to the connected local instances that cooperate and exchange data with the central system of the malware detection and protection system.
[0053] The term "auto-encoder" may refer to a specific type of artificial neural network that is used to learn data coding efficiently in an unsupervised manner. Thereby, the goal of an auto-encoder is to learn a representation (i.e., encoding) of a set of data, typically by training a neural network to ignore the "noise" in the signal, in order to reduce its dimensionality. Alongside the reduction side, a reconstruction side is also trained (i.e., learned), where the auto-encoder attempts to generate a representation from the reduced encoding that is as close as possible to the original input.
[0054] The term "regular process" may refer to an operating system process that is launched and / or controlled by the operating system, such as an internal process of the operating system or a process intended for an application.
[0055] The term "loss functions," in its simplest form, may represent the difference between output values and input values (e.g., from a neural network system, an autoencoder, etc.). Embodiments of the present invention recognize that loss functions are widely used in artificial intelligence (AI) systems.
[0056] Embodiments of the present invention provide a method for protecting a host (i.e., a computer system running a process) from the harmful effects of malware that arrives on the host through a network connection (e.g., as an email attachment, a program download, an app or applet within a web browser, etc.).
[0057] Embodiments of the present invention may utilize local sandboxing of unknown processes for anomaly detection. Incoming network traffic may first be subject to an initial check for indicators of compromise (IoCs), such as known malicious IP addresses or virus signatures. If the process is neither a known good nor known bad process, embodiments of the present invention may allow the process to run, but may initially be kept in a secure enclave, or sandbox.
[0058] The concept also includes a decision-making method (running locally on the host) to identify whether a process is malicious or not. Malicious processes are discarded and non-malicious processes are removed from the sandbox and run as normal processes.
[0059] In various embodiments, each local host runs two anomaly detectors or systems: the first one uses a federated machine learning model built from process telemetry data contributed by telemetry extractors running on all hosts, sharing the basic concepts of the inventive idea proposed here, and the second one uses a machine learning model built from locally extracted process telemetry data. In an example embodiment, the anomaly detector is an autoencoder. However, other machine learning implementations are possible. The machine learning model (MLM) is trained by telemetry data from normal (non-anomalous) processes (i.e., it describes the range of characteristics of normal process telemetry data). Deviations from this range are considered anomalous.
[0060] Embodiments of the present invention recognize that a challenge with this approach is to control the false positive rate (i.e., the relative amount of instances where a normal process is incorrectly classified as malware) while preventing malware from getting through as a normal process. Both a high false positive rate and malware getting through can have a significant impact on the effectiveness of a security system. The dual anomaly detector method allows for the introduction of a sensitivity factor to tune and optimize the true positive / false positive rate.
[0061] Various embodiments of the present invention recognize that for many years, cybersecurity attacks continue to be a top priority for government agencies and information technology (IT) organizations and companies. The role of a Chief Information Security Officer (CISO) is known in any IT organization. Embodiments of the present invention recognize that the majority of security information and event monitoring (SIEM) solutions available in the market require a complex set of correlation rules monitored and tuned by highly trained personnel to identify and detect potential security threats that may lead to security incidents in at least a portion of the organization. According to the General Data Security Regulation (GDPR), companies operating within the European Union may be subject to high fees if the company fails to protect customer and / or employee data. However, protecting corporate data is not just a topic for European-based companies and government agencies, but for all businesses that handle data around the world.
[0062] Additional embodiments of the present invention recognize that advanced malware may also detect centralized sandbox execution and may disguise malware behavior as a legitimate process. Thus, in developing a centralized sandbox system for malware analysis and unknown files / processes, embodiments of the present invention recognize that there is a high risk of bottlenecks. Embodiments of the present invention also recognize that the amount of available system resources required in exchange for the number of secure enclave executions performed by the centralized sandbox system may be negligible.
[0063] Further embodiments of the present invention recognize that a problem with using centralized sandboxing, however, is that it may require excessive resources of the host system. Thus, embodiments of the present invention recognize that there is a need to overcome this bottleneck and provide a method for (i) identifying malicious processes on a host that may not be detected by traditional malware detection systems, and (ii) requiring significantly fewer resources than centralized sandboxing approaches.
[0064] Additional embodiments applicable to the present method and related systems are described below.
[0065] FIG. 1 illustrates a block diagram of a preferred embodiment of a method 100 for detecting malicious (e.g., particularly anomalous or non-regular) processes by selected instances (i.e., hosts) of an anti-malware system, according to an embodiment of the present invention. For example, an anti-malware system including multiple hosts and a central system (e.g., cloud-based) may be used from time to time as needed without becoming a bottleneck. For illustrative purposes, the method 100 described in FIG. 1 may be implemented in one or more of the diagrams and systems illustrated in FIG. 6 (e.g., utilizing component 600), FIG. 7 (e.g., utilizing embodiment 700), FIG. 8 (e.g., utilizing sub-process 800), FIG. 9 (e.g., utilizing malware detection system 900), and FIG. 10 (e.g., utilizing computer system 1000). Alternatively, the execution of the method 100 is not limited to this implementation. In an exemplary embodiment, the malware detection system 900 may execute the method 100 of FIG. 1.
[0066] Method 100 includes examining a process for indicators of compromise (step 102). In an exemplary embodiment, the process relates to previously received program code or data. Method 100 also includes classifying the process based on the results of the examination (step 104). In various embodiments, method 100 may classify the process as one selected from the group consisting of a known good process, a known malicious process, and an unknown process.
[0067] At step 106, the method 100 includes executing the process in a secure enclave. If the process is not classified as either a known good process or a known malicious process, the method 100 executes the process in a secure enclave (i.e., a secure sandbox on the underlying host computing system). Additionally, the method 100 may operate to determine whether the process is deemed malicious. At step 108, the method 100 includes collecting telemetry data. In particular, static and dynamic telemetry data from the execution of the process under observation (running within the secure enclave).
[0068] Further, in step 110, method 100 includes passing the collected telemetry data to a locally trained neural network system. In various embodiments, training data for the locally trained neural network system (e.g., for developing a trained neural network model) includes telemetry data from processes (e.g., non-malicious processes, such as normal processes) executing on the host system underlying the local instance.
[0069] Additionally, in step 112, method 100 also includes passing the collected telemetry data to a federated trained neural network system. In an example embodiment, the federated trained neural network system is adapted to receive a federated trained neural network model, for example, from a supporting cloud system that is part of the malware detection system.
[0070] Further, in step 114, method 100 includes determining a result of a first loss function (e.g., mean square of two vectors) for the locally trained neural network system. Method 100 may then compare the result to a result of the loss function at the end of training of the locally trained neural network model. Then, in step 116, method 100 includes determining a result of a second loss function (e.g., mean square based) for the federated trained neural network system. Method 100 may then compare the result to a result of the loss function of the received federated trained neural network model of the federated trained neural network system.
[0071] Next, method 100 includes aggregating results of the first and second loss functions (step 118). Further, method 100 includes determining whether the process is anomalous based on the aggregated results (step 120). In various embodiments, the aggregation may use a predefined process to determine the aggregated results (see below for examples described in more detail).
[0072] Figure 2 shows a diagram 200 illustrating telemetry data of a non-anomalous process. Figure 2 shows the distribution of input telemetry data for the machine learning models (MLM) of the federated anomaly detector (broader distribution on the left: federated trained neural network model) and the local anomaly detector (narrower distribution 204: locally trained neural network system). The x-axis of diagram 200 shows categories of telemetry data characteristics, while the y-axis of diagram 200 shows the relative frequency of their occurrence.
[0073] In the illustrated example of FIG. 2, the closer the characteristics of the telemetry data for the processes currently under investigation (specifically, Process A (206), Process B (208) and Process C (210)) are to the center of the distribution (i.e., the apex of the Gaussian curve), the higher the confidence that the process is normal (i.e., non-malicious).
[0074] An embodiment of the present invention recognizes that MLM does not represent the real world, but instead represents a world based on input data, which is always somewhat biased. In this example, the characteristics of the telemetry data collected on this particular local host are different from the characteristics collected across the environment. Process B 208 is in fact a normal process on the local host, even though Process B is not considered as a normal process by the federated anomaly detector due to different training data. In order to avoid blocking a normal process on a system with different characteristics, in principle both anomaly detectors should classify the process as anomalous, which is the case for Process C (210). In the example of FIG. 2, Process C (210) is outside the range of Gaussian curve 202 and Gaussian curve 204. Using model-specific sensitivity parameters and weighting the results within the anomaly detectors (i.e., each trained neural network model / system), an embodiment of the present invention can operate to improve the true / false positive rate.
[0075] FIG. 3 shows the structure of an autoencoder 300. In the illustrated example, the autoencoder 300 is shown in a typical implementation of a neural network that includes multiple layers of nodes. As mentioned above, an autoencoder is a type of artificial neural network that is used to learn effective data coding in an unsupervised manner. The goal of an autoencoder is to learn a representation (i.e., encoding) for a set of data, typically by training the network to ignore the signal "noise" for dimensionality reduction. In conjunction with the reduction side, a reconstruction side is trained, where the autoencoder attempts to generate a representation from the reduced encoding that is as close as possible to the original input, hence the name. Hence, after training, the output node 306 provides the same data signal that was input to the input layer 302 in the inference phase. In the example shown in FIG. 3, each circle represents a node of the neural network or a cell of the artificial neural network. The autoencoder 300 is shown with only three layers 302, 304, 306, where layer 304 is a hidden layer. In general, an autoencoder may contain multiple hidden layers, however the number of layers does not change the general functionality of the autoencoder.
[0076] FIG. 4 shows a first part, flowchart 400, of a more detailed exemplary flowchart of an embodiment of the present invention. First, inbound network traffic 402 undergoes traffic analysis 404. Here, an embodiment of the present invention utilizes flowchart 400 to make a three-pronged decision of whether the received data belongs to known benign / allow-listed traffic 406 (left branch), known malicious traffic 408 (left branch) that can be blocked immediately, or unknown traffic 410 that leaves traffic analysis 404 down. Unknown traffic cannot be directly classified as benign or bad (in unknown traffic 410). The unknown traffic may then be fed to sub-process 412, "Network Analysis Engine Process Workflow," and (via packet cache) sub-process 414, "Host Analysis Engine Process Workflow." Additionally, if the deep packet analysis of sub-process 412 identifies non-malicious traffic, then the information is forwarded to host analysis engine process workflow 414. The process flow continues in FIG. 5. The flow charts of Figures 4 and 5 are linked by connections "A", "B" and "C".
[0077] 5 illustrates a second portion of a more detailed exemplary flowchart of an embodiment of the present invention, flowchart 500. The process flow continues from FIG. 4 at "A." The hosted analysis engine process workflow 414 (compare FIG. 4) enables process execution in a secure enclave (i.e., a secure sandbox execution environment) in process 502 after confirmation from the network analysis engine (FIG. 4, process 412).
[0078] Flowchart 500 then begins process 504, recording of telemetry data, while embodiments of the invention may perform process 506, extraction of metadata from the telemetry data and generated alerts. Flowchart 500 continues with process 508, determining whether an anomaly is detected. If an anomaly is detected (process 508, Y branch), the process is aborted (at process 510) and the associated data is deleted. Additionally, flowchart 500 includes process 512, forwarding the collected metadata in updated form to the host and network engine, and the analysis process returns to FIG. 4 along connections "B" and "C".
[0079] If the determination in process 508 indicates that no anomalies are detected (i.e., no malicious code / data) (process 508, N branch), flowchart 500 proceeds to process 514, where the process runs in the normal manner on the host. Security monitoring for the particular process is then terminated (process 516).
[0080] 6 illustrates a block diagram of an embodiment of components 600 of various aspects of the invention. In response to executing an unclassifiable process (enclave process execution 602) as a secure enclave process, a telemetry recorder and metadata extractor 604 collects data generated from the enclave process execution 602. First, a telemetry recorder 606 collects the data and passes the data to a telemetry and metadata extractor module 608. The telemetry recorder and metadata extractor 604 may then forward the data to an anomaly detector 610, from which the data may be forwarded (via forwarding) to a metadata updater 614.
[0081] Once the data has been transferred to the metadata updater 614, the process is automatically abandoned and any associated process output data is deleted. Otherwise, the host process executor 612 is started. However, the host process executor 612 is only activated if the anomaly detector 610 identifies a non-malicious process.
[0082] FIG. 7 illustrates a block diagram of an embodiment 700 illustrating a process of generating telemetry and metadata for a given process to be inspected according to various embodiments of the present invention. A telemetry recorder 702 can receive data from a process / file extractor 701. The telemetry recorder 702 operates in two steps to record executable processes and other related data using a static telemetry (TM) recorder 704 and a dynamic telemetry recorder 706. The static telemetry recorder 704 records content obtained directly before process / code execution, which may include (i) a determined file hash value of the executable file for fingerprinting the file (e.g., file fingerprinting 708), (ii) code string extraction from the executable file for code string analysis 710, (iii) executable layout extraction 712 of the executable file, and (iv) automatic disassembly 714 of the executable file's code and code structure.
[0083] In addition, the dynamic telemetry recorder 706 records the capturing / code execution of the obtained content, which includes at least capturing and recording (i) system monitoring data, (ii) registry activity data 716, (iii) file activity data 718 (i.e., file access), (iv) process activity data 720, and (v) network activity data 722.
[0084] As a result, the telemetry (TM) and metadata extractor 724 can provide static telemetry extraction data 726 including fingerprint hash data, strings in code, executable layout / patterns of program code of executable files, and disassembly information. In other aspects, the dynamic telemetry extraction data 728 includes exemplary registry dumps, file activity extractions, process dumps, and network dumps. For both, the static telemetry recorder 704 and the dynamic telemetry recorder 706 may be partially reused (and existing technology, if applicable).
[0085] 8 shows a block diagram of an embodiment illustrating a sub-process 800 for anomaly detection including the use of a trained neural network module, according to various embodiments of the present invention. In various embodiments, the main components involved in the execution of the sub-process 8 include a local anomaly detector 804 (i.e., a locally trained neural network system), a federated anomaly detector 806 (i.e., a federated trained neural network system), a training data collector 812, a retraining module 814, a multi-set model federated learning unit 820, and a provided anomaly detection model (e.g., provided as a cloud service such as an anomaly detection model cloud 818), with the last two components being part of a central service or system 816.
[0086] The stream of telemetry data enters a temporal aggregator 802. For example, the temporal aggregator 802 aggregates the data stream for n seconds before sending the data to the next module, a local anomaly detector 804 (e.g., in the form of an autoencoder). In parallel, the aggregated telemetry data is parsed to a federated anomaly detector 806 (e.g., also in the form of an autoencoder). In block 808, both results are aggregated to form anomaly estimations. A decision module 810 decides whether the "anomaly true" or "anomaly false" condition is met.
[0087] The telemetry data may then proceed (e.g., via transfer) to the training data collector 812, which then passes the telemetry data to the retraining module 814. The retraining module 814 can serve to train the locally trained neural network model (i.e., the local anomaly detector 804) periodically (e.g., daily, hourly, or according to any other timing scheme). Alternatively, the retraining may be initiated by a predefined trigger mechanism. Furthermore, the retraining module 814 pulses the telemetry data to a central system 816, which may include a (cloud-based) federated anomaly detector or anomaly detection model cloud 818. In addition, the anomaly detection model cloud 818 may be retrained using a multiple-aggregated model federated learning unit 820 based on multiple aggregated neural network models from different hosts from which the locally trained neural network model is used.
[0088] Thus, various embodiments of the present invention recognize that there are advantages to having both federated and local anomaly detectors, not only in terms of resource consumption, but also in terms of the data used for training. In effect, the federated learning autoencoder is exposed to data from all instances of anti-malware systems around the environment, and the federated learning autoencoder learns on the data without directly transferring any information to the autoencoder of any single instance. The transfer of knowledge of the different instances is closed within the federated trained neural network model, which is then periodically transferred to the neural network system of the federated anomaly detector in the local host.
[0089] During training, the autoencoder is exposed to aggregated telemetry data as input and output, and the autoencoder learns to reproduce the aggregated telemetry data (because of the autoencoder functionality). A typical loss function used is the mean squared error function (MSE).
[0090] During inference (i.e., the use of the model in the process shown in Figure 8), the local and federated autoencoders are used with the current telemetry data extracted from the unknown code running in the secure enclave. The same MSE metric is then calculated independently for both autoencoders between the reconstructed data and the original data. The values are then aggregated as follows: for both autoencoders (AE), the result is m AE =MSE current -(MSE current - during training) / sf, where "sf" is the sensitivity factor.
[0091] If the resulting value is negative for AE, no anomaly is detected. Otherwise, the margin m AE is the weighted average, e.g., m AE +W fed *m AE,fed The weights Wfed may depend on the amount of data seen by the federated AEs versus the local AEs. In one embodiment, a linear dependency is used so that federated AEs that learn from more data have higher weight values. As a fourth step, m AE,loc+ W fed *m AE,fed If it is positive, an anomaly is identified and the process under test is discarded, otherwise, as described above, the process is not anomalous and the telemetry data is passed to the training data collector.
[0092] It should also be mentioned that the training data collector stores all telemetry data that passes through the detector and is classified as normal. Periodically, the training data collector sends a badge of the data to the retraining module.
[0093] The retraining module uses incoming telemetry data from the training data collector to retire the training of the AE. For example, a typical period is one day. Various embodiments of the present invention can automatically determine when training is complete using a condition based on the improvement of the performance of the neural network model during training. In one example, a suitable condition is early stopping. Once model training is complete, the model of the local AE is replaced with a new model, and the new model is sent to the federated learning module. The multi-set model federated learning, running at a central computing location (e.g., as a cloud service), periodically receives a set of new models (with a configurable time range) and integrates the incoming models to ensure that the single local neural network model (or the knowledge of the local anomaly detector) is not forgotten.
[0094] In various aspects, to ensure accurate utilization of the single local neural network model (or the knowledge of the local anomaly detector), the process includes the following: First, a set of typically generated telemetry data is processed through each single collected locally trained neural network system and the other side is stored. The output will encode the experience of the single machine learning model. Second, the pairs (AE input / output) generated by the single model are obtained and a federated neural network model is trained with these data pairs. Once training is complete, the new federated trained neural network model is sent to the anomaly detection model cloud 818, which then pushes the new federated trained neural network model to all the federated anomaly detectors on the local host. The federated anomaly detectors can receive the new model and replace the model at the appropriate moment.
[0095] In one exemplary embodiment, a single value for the sensitivity factor of the federated AE is assigned based on the true / false positive rates observed by the security analyst. In a modified exemplary embodiment, the sensitivity factor sf on the local AE is loc is adjusted based on a combination of the following automatically observed parameters: The first parameter may be the amount of local telemetry measurement processes to build a local model (e.g., for a relatively new server, only a few samples describing normal behavior are incorporated into the model, which may easily trigger a false alarm (detecting a malicious process from malware)). For example, the sensitivity factor can be set to a very low value close to zero.
[0096] The second parameter may be the degree of variation of the local telemetry measurement process to build the local model. For example, a server with a specific objective may receive a specific set of inputs that describe normal behavior, resulting in a biased model. A lower sf may be used to avoid false positive alarms. loc is assigned to a small variance σ. Servers that receive a wider set of telemetry data will build more balanced models. Here, for large variance σ, values close to 1 are assigned to sf loc is assigned to.
[0097] The variance of the telemetry input data can be defined as follows:
number
number
[0098] The third parameter may be the amount of overlap between the federated model and the local model (compare Figure 2). The overlap is the sum of the results across the cases studied. AE,loc -m AE,fed For large overlaps, the sensitivity factor sf close to zero loc is assigned (hosts are common hosts), there is a small overlap (especially in combination with the small variance of the telemetry data measurements), and a low value for the sensitivity factor sf loc is assigned to.
[0099] 9 illustrates a block diagram of an embodiment of a proposed malware detection system 900 according to various aspects of the present invention. The malware detection system 900 includes a memory and a processor 902 communicatively coupled to the memory 904, the processor 902 configured to use program code stored in the memory 904 to inspect processes for indicators of compromise (particularly by an inspection unit 906) and to classify the process as one selected from the group consisting of a known good process, a known malicious process, and an unknown process based on the results of the inspection unit 906 (particularly by a classification module 908).
[0100] The processor 902 is further configured to use program code stored in the memory 904 to execute a process within a secure enclave (particularly by the execution unit 910) and collect (particularly by the collection module 912) telemetry data from the executing process if the process is not classified as either a known good process or a known malicious process.
[0101] Further, the processor 902 is also configured, using program code stored in the memory 904, to pass the collected telemetry data (particularly by the first passing unit 914) to the locally trained neural network system, and to pass the collected telemetry data (particularly by the second passing unit 916) to the federated trained neural network system, where the training data of the locally trained neural network system includes telemetry data from a process running on a host system underlying the local instance, and where the federated trained neural network system is adapted to receive the federated trained neural network model.
[0102] In addition, the processor 902 is configured to use program code stored in the memory 904 to determine (particularly, by the first determination module 918) a result of a first loss function for the locally trained neural network system and compare the result with a result of the loss function at the end of training of the locally trained neural network system, and to determine (particularly, by the second determination module 920) a result of a second loss function for the federated trained neural network system and compare the result with a result of the loss function of the received federated trained neural network model of the federated trained neural network system.
[0103] Last but not least, the processor 902 is configured, using program code stored in the memory 904, to aggregate (particularly, by an aggregation module 922) results of the first and second loss functions and to determine (particularly, by a third decision module 924) based on the aggregated results whether the process is anomalous or not.
[0104] It should also be mentioned that all units and modules may be implemented as hardware units and modules, and the units are communicatively coupled to each other for signal and data exchange. Alternatively, for a one-to-one communication scheme, the units and modules may also be connected to the internal bus system 926 of the malware detection system (i.e., the processor 902, the memory 904, the inspection unit 906, the classification module 908, the execution unit 910, the collection module 912, the first passing unit 914, the second passing unit 916, the first decision module 918, the second decision module 920, the aggregation module 922 and the third decision module may all be connected to the internal bus system 926 of the malware detection system for signal and data exchange).
[0105] The embodiments of the present invention may be implemented using substantially any type of computer, regardless of whether the platform is suitable for storing and / or executing program code. Figure 10 shows, as an example, a computer system 1000 suitable for executing program code related to the proposed method according to various embodiments of the present invention.
[0106] Computer system 1000 is one example of a suitable computer system, and whether computer system 1000 is capable of implementing and / or performing any of the functionality described above is not intended to suggest any limitation on the scope of use or functionality of the embodiments of the invention described herein. In computer system 1000, there are components that are operable in numerous other general purpose or special purpose computing system environments or configurations. Examples of well-known computing systems, environments and / or configurations suitable for use with computer system 1000 include, but are not limited to, personal computer systems, server computer systems, thin clients, thick clients, handheld or laptop devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputer systems, mainframe computer systems, distributed cloud computing environments including any of the above-mentioned systems or devices, and the like. Computer system 1000 (e.g., a server) may be described in the general context of computer system-executable instructions, such as program modules, executed by computer system 1000. Generally, program modules may include routines, programs, objects, components, logic, data structures, etc. that perform particular tasks or implement particular abstract data types. The computer system 1000 (e.g., a server) may be implemented in a distributed cloud computing environment where tasks may be performed by remote processing devices linked through a communications network. In a distributed cloud computing environment, program modules may be located in both local and remote computer system storage media, including memory storage devices.
[0107] As shown in FIG. 10, a computer system 1000 (e.g., a server) is shown in the form of a general-purpose computing device. Components of the computer system 1000 (e.g., a server) include, but are not limited to, one or more processors or processing units 1002, a system memory 1004, and a bus 1006 that couples various system components, including the system memory 1004, to the processing unit 1002. The bus 1006 may represent one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, and a processor or local bus using any of a variety of bus architectures. By way of example and without limitation, such architectures include an Industry Standard Architecture (ISA) bus, a MicroChannel Architecture (MCA) bus, an Enhanced ISA (EISA) bus, a Video Electronics Standards Association (VESA) local bus, and a Peripheral Component Interconnect (PCI) bus. The computer system 1000 (e.g., a server) may typically include a variety of computer system readable media. Such media may be any available media that is accessible by the computer system 1000 (eg, a server) and includes both volatile and non-volatile media, removable and non-removable media.
[0108] The system memory 1004 may include computer system readable media in the form of volatile memory such as random access memory (RAM) 1008 and / or cache memory 1010. The computer system 1000 (e.g., a server) may further include other removable / non-removable volatile / non-volatile computer system storage media. By way of example only, the storage system 1012 may be provided for reading from and writing to a non-removable non-volatile magnetic medium (not shown, typically referred to as a "hard drive"). Although not shown, a magnetic disk drive may be provided for reading from and writing to a removable non-volatile magnetic disk (e.g., a floppy disk), as well as an optical disk drive for reading from and writing to a removable non-volatile optical disk, such as a CD-ROM, DVD-ROM, or other optical medium. In such an example, each may be connected to the bus 1006 by one or more data media interfaces. As further shown and described, the memory 1004 may include at least one program product having a set (eg, at least one) of program modules configured to perform functions of embodiments of the present invention.
[0109] The programs / utilities include a set (at least one) of program modules 1016, which may be stored in memory 1004, by way of example only, including, but not limited to, an operating system, one or more application programs, other program modules, and program data. Each of the operating system, one or more application programs, other program modules, and program data, or some combination thereof, may include an implementation of a networking environment. The program modules 1016 may perform the functions and / or methodologies of embodiments of the present invention as described herein.
[0110] The computer system 1000 (e.g., a server) may also communicate with one or more external devices 1018, such as a keyboard, a pointing device, a display 1020, one or more devices that allow a user to interact with the computer system 1000 (e.g., a server), and / or any device (e.g., a network card, a modem, etc.) that allows the computer system 1000 (e.g., a server) to communicate with one or more other computing devices. Such communication may occur through an input / output (I / O) interface 1014. Additionally, the computer system 1000 (e.g., a server) may communicate with one or more networks, such as a local area network (LAN), a general wide area network (WAN), and / or a public network (e.g., the Internet), through a network adapter 1022. As shown, the network adapter 1022 may communicate with other components of the computer system 1000 (e.g., a server) through a bus 1006. Although not shown, other hardware and / or software components may be used in combination with the computer system 1000 (e.g., a server). Examples may include, but are not limited to, microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data archive storage systems.
[0111] Additionally, a malware detection system 900 for detecting malicious processes by a selected instance of an anti-malware system may be attached to the bus 1006.
[0112] The programs described herein are identified based on the application for which they are implemented in a particular embodiment of the invention. However, it should be understood that any particular program nomenclature herein is used for convenience only, and thus the invention should not be limited to use in any particular application identified and / or implied by such nomenclature.
[0113] The present invention may be a system, method and / or computer program product at any possible level of technical detail of integration. The computer program product may include a computer-readable storage medium having computer-readable program instructions thereon for causing a processor to perform aspects of the present invention.
[0114] A computer readable storage medium may be a tangible device that holds and stores instructions for use by an instruction execution device. A computer readable storage medium may be, for example, but not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above. An illustrative enumeration of more specific examples of computer readable storage media includes portable computer diskettes, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROMs or flash memories), static random access memories (SRAMs), portable compact disk read-only memories (CD-ROMs), digital versatile disks (DVDs), memory sticks, floppy disks, punch cards or mechanically encoded devices such as ridge structures in grooves having instructions recorded thereon, and any suitable combination of the above. Computer-readable storage media, as used herein, is not to be construed as a transitory signal per se, such as an electric wave, a freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse passing through a fiber optic cable), or an electrical signal transmitted through a wire.
[0115] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to the respective computing / processing device or to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network, or a combination thereof. The network may include copper transmission cables, optical transmission fiber, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers, or a combination thereof. A network adapter card or network interface in the computer / processing device(s) receives the computer-readable program instructions from the network and transfers the computer-readable program instructions for storage in the computer-readable storage medium within the respective computing / processing device.
[0116] The computer readable program instructions for carrying out the operations of the present invention may be assembler instructions, instruction set architecture (ISA) instructions, machine language instructions, machine dependent instructions, microcode, firmware instructions, state setting data, configuration data for an integrated circuit, or source or object code written in any combination of one or more programming languages, including object oriented languages such as Smalltalk®, C++ or the like, procedural languages such as the C programming language or similar programming languages. The computer readable program instructions may be executed as a stand-alone software package, entirely on the user's computer, partially on the user's computer, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (e.g., through the Internet using an Internet Service Provider). In some embodiments, the electrical circuitry may execute the computer readable program instructions by utilizing state information of the computer readable program instructions to individualize the electrical circuitry, which may include, for example, a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), to perform aspects of the invention.
[0117] Aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, may be implemented by computer readable program instructions.
[0118] These computer readable program instructions may be provided to a computer processor or other programmable data processing apparatus to produce a machine such that the instructions, when executed via the computer processor or other programmable data processing apparatus, create means for implementing the functions / acts identified in a block or blocks of the flowchart illustrations and / or blocks. These computer readable program instructions may also be stored on a computer readable storage medium capable of directing a computer, programmable data processing apparatus and / or other device to function in a particular manner, such that a computer readable storage medium having instructions stored thereon includes an article of manufacture including instructions that implement aspects of the functions / acts identified in a block or blocks of the flowchart illustrations and / or blocks.
[0119] The computer readable program instructions may also be loaded into a computer, another programmable data processing apparatus, or other device and cause the computer, other programmable data processing apparatus, or other device to perform a series of operational steps to generate a computer-implemented process such that the instructions executing on the computer, other programmable data processing apparatus, or other device implement an aspect of the function / action identified in a block or blocks of the flowcharts and / or blocks.
[0120] The flowcharts and block diagrams in the drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowcharts or block diagrams may represent a module, segment, or portion of instructions, including one or more executable instructions for implementing a particular logical function(s). In some alternative implementations, the functions noted in the blocks may occur out of the order depicted in the drawings. For example, two blocks shown in succession may in fact be completed as one step, may be executed simultaneously, may be executed substantially simultaneously in a partially or fully overlapping manner, or the blocks may be executed in reverse order depending on the functionality involved. It should be noted that each block of both the block diagrams and / or flowchart diagrams, and combinations of blocks in the block diagrams and / or flowchart diagrams may be implemented by a special-purpose hardware-based system that performs a particular function or action, or that implements a combination of special-purpose hardware and computer instructions.
[0121] The description of various embodiments of the present invention is presented for illustrative purposes, but is not intended to be exhaustive or limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope of the present invention. The terms used in this specification are selected to best explain the principles of the embodiments, practical applications, or technical improvements to the technology found in the market, or to allow other skilled in the art to understand the embodiments disclosed herein.
Claims
1. examining, by one or more processors, a process for indicators of compromise to said process; determining, by one or more processors, a classification of the process based on results of the tests; and in response to determining that the classification of the process does not correspond to known good processes and known malicious processes, executing, by one or more processors, the process within a secure enclave on a local host system; collecting, by one or more processors, telemetry data from the processes executing within the secure enclave; passing, by one or more processors, the telemetry data collected from the secure enclave to a locally trained neural network system and a federated trained neural network system, where training data for the locally trained neural network system includes telemetry data from processes executing on the local host system that underlie the locally trained neural network system; determining, by one or more processors, a result of a first loss function for the locally trained neural network system; comparing, by one or more processors, the result with a result of the first loss function at the end of training of the locally trained neural network system; aggregating, by one or more processors, results of the first loss function and results of a second loss function for a neural network system that is federated trained separately from the locally trained neural network system; determining, by one or more processors, whether the process is abnormal based on the results of the aggregation; and A method comprising:
2. The associatively trained neural network system comprises: determining, by one or more processors, a result of a second loss function for the jointly trained neural network system; and comparing, by one or more processors, the result with a result of the second loss function of the received jointly trained neural network model of the jointly trained neural network system; The method of claim 1 further comprising:
3. collecting, by one or more processors, telemetry data of non-malicious processes executing on selected instances of the anti-malware system; retraining, by one or more processors, the locally trained neural network system using the collected telemetry data to periodically build an updated local neural network model. The method of claim 1 further comprising:
4. 3. The method of claim 2, further comprising receiving, by one or more processors, an updated federated-trained neural network model for the federated-trained neural network system, the federated-trained neural network model being constructed using the locally trained neural network systems of the selected instances as input.
5. The method of claim 2 , wherein at least one of the locally trained neural network system and the federated trained neural network system is an autoencoder system.
6. The method of claim 1 , further comprising: in response to determining that the classification of the process corresponds to an abnormal process, discarding, by one or more processors, the process.
7. in response to determining that the process is a normal process, migrating, by one or more processors, the process outside the secure enclave based on execution in the secure enclave; executing said process as a normal process by one or more processors; The method of claim 1 further comprising:
8. The method of claim 2 , wherein the federated trained neural network system is trained using telemetry data from multiple hosts.
9. Training the associatively trained neural network system includes: processing, by one or more processors, the set of representative telemetry data through each of a plurality of received locally trained neural network models and storing the respective results; training, by one or more processors, the associatively trained neural network model using telemetry data input / output pairs used and generated during processing of the telemetry data as input data for training the associatively trained neural network model; The method of claim 8 further comprising:
10. the telemetry data input / output pairs used to train the federated trained neural network model are weighted according to the geographical proximity between the geographical host location of the locally received neural network model and the geographical host location at which a new federated trained neural network model is trained; 10. The method of claim 9, wherein the telemetry data input / output pairs used to train the federated trained neural network model are weighted depending on a predefined metric.
11. Aggregating results of the first loss function and the second loss function includes: constructing, by one or more processors, a weighted average of the first loss function and the second loss function. The method of claim 2 , further comprising:
12. The method of claim 1 , wherein the determined classification is selected from the group consisting of a known good process, a known malicious process, and an unknown process.
13. 1. A computer system comprising: one or more computer processors; one or more computer readable storage media; program instructions stored in a computer-readable storage medium for execution by at least one of the one or more processors; the program instructions comprising: program instructions for examining a process for indicators of infringement for said process; program instructions for determining a classification of the process based on results of the inspection; in response to determining that the classification of the process does not correspond to a known good process and a known malicious process, program instructions for executing the process within a secure enclave on a local host computer; program instructions for collecting telemetry data from the process executing within the secure enclave; and passing the telemetry data collected from the secure enclave to a locally trained neural network process and a federated training neural network process, wherein training data for the locally trained neural network process includes telemetry data from processes executing on the local host computer that underlie the locally trained neural network process; program instructions for determining a result of a first loss function for the locally trained neural network process; program instructions for comparing the result with a result of the first loss function at the end of training of the locally trained neural network system; program instructions for aggregating results of the first loss function and results of a second loss function for a neural network process that is federated trained separately from the locally trained neural network process; program instructions for determining whether the process is abnormal based on the results of the aggregation; 2. A computer system comprising:
14. the associatively trained neural network process is adapted to receive an associatively trained neural network model; determining a result of a second loss function for the associatively trained neural network system; and comparing the result with a result of the second loss function of the received jointly trained neural network model of the jointly trained neural network system; 14. The computer system of claim 13, further comprising program instructions stored on the computer readable storage medium for execution by at least one of the one or more processors.
15. collecting telemetry data of non-malicious processes running on a selected instance of the anti-malware system; retraining the locally trained neural network system using the collected telemetry data to periodically build an updated local neural network model; 14. The computer system of claim 13, further comprising program instructions stored on a computer readable storage medium for execution by at least one of the one or more processors to:
16. 14. The computer system of claim 13, further comprising program instructions stored in a computer readable storage medium for execution by at least one of the one or more processors to discard the process in response to determining that the classification of the process corresponds to an anomalous process.
17. in response to determining that the process is a normal process, migrating the process outside the secure enclave based on execution in the secure enclave; running said process as a normal process; 14. The computer system of claim 13, further comprising program instructions stored in a computer readable storage medium for execution by at least one of the one or more processors to:
18. 15. The computer system of claim 14, wherein the federated trained neural network system is trained using telemetry data from multiple hosts.
19. A computer program comprising: one or more computer readable storage media; and program instructions stored on the one or more computer readable storage media, the program instructions comprising: program instructions for examining a process for indicators of infringement for said process; program instructions for determining a classification of the process based on results of the inspection; in response to determining that the classification of the process does not correspond to a known good process and a known malicious process, program instructions for executing the process within a secure enclave on a local host computer; program instructions for collecting telemetry data from the process executing within the secure enclave; and passing the telemetry data collected from the secure enclave to a locally trained neural network process and a federated training neural network process, wherein training data for the locally trained neural network process includes telemetry data from processes executing on the local host computer that underlie the locally trained neural network process; program instructions for determining a result of a first loss function for the locally trained neural network process; program instructions for comparing the result with a result of the first loss function at the end of training of the locally trained neural network system; program instructions for aggregating results of the first loss function and results of a second loss function for a neural network process that is federated trained separately from the locally trained neural network process; program instructions for determining whether the process is abnormal based on the results of the aggregation; A computer program comprising:
20. the associatively trained neural network process is adapted to receive an associatively trained neural network model; determining a result of a second loss function for the associatively trained neural network system; and comparing the result with a result of the second loss function of the received jointly trained neural network model of the jointly trained neural network system; 20. The computer program product of claim 19, further comprising program instructions stored on the computer readable storage medium for causing the computer program product to:
Citation Information
Patent Citations
Electronic message analysis for malware detection
JP2014513834A
Collaborative and Adaptive Threat Intelligence for Computer Security
US20150373043A1
Deployment of Machine Learning Models for Discernment of Threats
US20170357807A1
Malicious object detection in a runtime environment
US20180189489A1
Methods and apparatus for federated training of a neural network using trusted edge devices
US20190042937A1
