WO2020143227A1 - 一种基于对抗学习的工业控制系统恶意样本生成方法 - Google Patents

一种基于对抗学习的工业控制系统恶意样本生成方法 Download PDF

Info

Publication number
WO2020143227A1
WO2020143227A1 PCT/CN2019/101247 CN2019101247W WO2020143227A1 WO 2020143227 A1 WO2020143227 A1 WO 2020143227A1 CN 2019101247 W CN2019101247 W CN 2019101247W WO 2020143227 A1 WO2020143227 A1 WO 2020143227A1
Authority
WO
WIPO (PCT)
Prior art keywords
industrial control
sample
adversarial
control system
learning
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2019/101247
Other languages
English (en)
French (fr)
Inventor
程鹏
何阳
陈积明
王文海
孙优贤
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University
Original Assignee
Zhejiang University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang University filed Critical Zhejiang University
Priority to US16/982,056 priority Critical patent/US20210319113A1/en
Priority to JP2020540726A priority patent/JP6932270B2/ja
Publication of WO2020143227A1 publication Critical patent/WO2020143227A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
    • G06F21/50—Monitoring users, programs or devices to maintain the integrity of platforms, e.g. of processors, firmware or operating systems
    • G06F21/57—Certifying or maintaining trusted computer platforms, e.g. secure boots or power-downs, version controls, system software checks, secure updates or assessing vulnerabilities
    • G06F21/577—Assessing vulnerabilities and evaluating computer system security
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/0265—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion
    • G05B13/027—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion using neural networks only
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B19/00—Program-control systems
    • G05B19/02—Program-control systems electric
    • G05B19/418—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM]
    • G05B19/4185—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM] characterised by the network communication
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
    • G06F21/50—Monitoring users, programs or devices to maintain the integrity of platforms, e.g. of processors, firmware or operating systems
    • G06F21/55—Detecting local intrusion or implementing counter-measures
    • G06F21/56—Computer malware detection or handling, e.g. anti-virus arrangements
    • G06F21/566—Dynamic detection, i.e. detection performed at run-time, e.g. emulation, suspicious activities
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0499—Feedforward networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/094—Adversarial learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/096—Transfer learning
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B2219/00—Program-control systems
    • G05B2219/30—Nc systems
    • G05B2219/31—From computer integrated manufacturing till monitoring
    • G05B2219/31244—Safety, reconnect network automatically if broken
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2221/00—Indexing scheme relating to security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
    • G06F2221/03—Indexing scheme relating to G06F21/50, monitoring users, programs or devices to maintain the integrity of platforms
    • G06F2221/034—Test or assess a computer or a system
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/02—Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]

Definitions

  • the invention belongs to the field of industrial control system security, and relates to a method for generating malicious samples of an industrial control system based on adversarial learning.
  • the intrusion detectors of existing industrial control systems based on machine learning algorithms are vulnerable.
  • the machine learning model is easily manipulated by the adversarial samples intentionally generated by the attacker when it is actually applied.
  • the attacker can make the algorithm model produce completely different results from the normal output by making minor modifications to the test examples.
  • the samples that make the model produce incorrect predictions are called adversarial samples.
  • Research that actively generates adversarial attacks on machine learning designs is called adversarial machine learning research. This counterattack brings potential security threats to systems using machine learning, especially for systems with high security requirements, such as industrial control.
  • the purpose of the present invention is to provide a method for generating a malicious sample of an industrial control system based on adversarial learning in view of the security deficiencies and deficiencies of the current industrial control system.
  • a method for generating a malicious sample of an industrial control system based on adversarial learning including the following steps:
  • the anti-sample producers sniff the communication data of the industrial control system, obtain the communication data distributed with the training data used by the industrial control intrusion detection system, and mark the communication data with a category label, the category includes abnormal and normal, where The abnormal communication data is used as the original attack sample.
  • the industrial control intrusion detection system is an existing industrial control intrusion detection system based on a machine learning method.
  • step (3) Establish a machine learning classifier based on the effective features extracted in step (2), train the machine learning classifier using the labeled industrial control system communication data, and obtain a trained classifier for distinguishing normal or abnormal communication data .
  • step (3) Using the classifier established in step (3), transform the adversarial learning problem of the industrial control intrusion detection system into the following optimization problem, and solve the optimization problem to obtain the final adversarial sample.
  • g(x) is the probability that the adversarial sample x * is judged to be an abnormal sample, calculated by the classifier
  • d(x * , x 0 ) is the distance between the adversarial sample and the original attack sample
  • d max indicates that the industrial control system allows The maximum Euclidean distance is that exceeding this distance has no malicious effect against the sample.
  • step (4) Test the countermeasure sample generated in step (4) in the actual industrial control system. If it can successfully bypass the industrial control intrusion detection system and retain the attack effect, it is regarded as an effective countermeasure sample, otherwise the countermeasure sample is discarded.
  • the producer of the confrontation sample should be a black box attacker, and cannot directly obtain the same data as the industrial control intrusion detection system (detector).
  • step (2) different effective features can be extracted for different industrial control system communication protocols.
  • Commonly used industrial control system communication protocols include Modbus, PROFIBUS, DNP3, BACnet, Siemens S7, etc.
  • Each communication protocol has a corresponding format and application scenario, and different communication protocols can be parsed according to the specific scenario to obtain an effective feature set.
  • the classifier used for the training of the adversarial sample generator may be a classifier different from that of the industrial control intrusion detection system, that is, the detection party.
  • the classifier generated by the adversarial sample generator may be called an adversarial
  • the principle of the local substitute model of learning is the mitigation of learning attacks.
  • step (4) specific optimization problem solving methods include gradient descent method, Newton method, COBYLA (for Constrained Optimization, BYLinear Approximations) method, etc.
  • the expression method of the distance includes a first normal distance, a second normal distance, and an infinite normal distance.
  • the machine learning classifier uses a neural network, and its probability calculation method is as follows:
  • x (i) is the i-th feature of the sample x
  • y is the label j corresponding to the sample x
  • ⁇ is the neural network parameter
  • ⁇ j is the neural network parameter corresponding to the label j
  • k is the label total
  • step (4) for a specific control scenario, special constraints on variables need to be added to the optimization problem. If the package function code of a scenario is determined, the adversarial sample cannot be changed in this dimension. That is, when this method is applied, in order to ensure that the generated adversarial samples can effectively complete the malicious attack effect, the generator needs to add different constraints to the design of the optimization variables of specific dimensions according to specific scenarios.
  • the beneficial effect of the present invention is: a method for generating malicious samples of an industrial control system based on adversarial learning, which increases the security of the industrial control system.
  • the original research field of adversarial learning has been expanded from machine vision and speech to the field of industrial control.
  • the security performance of industrial control intrusion detection system based on machine learning is enhanced to prevent attacks from malicious samples.
  • Figure 1 is a block diagram of the method of the present invention.
  • Fig. 2 is a simulation result diagram of the sub-invention method.
  • the method for generating malicious samples of an industrial control system based on confrontation learning provided by the present invention includes the following steps:
  • the anti-sample producers sniff the communication data of the industrial control system, obtain the communication data distributed with the training data used by the industrial control intrusion detection system, and mark the communication data with a category label, the category includes abnormal and normal, where The abnormal communication data is used as the original attack sample.
  • the industrial control intrusion detection system is an existing industrial control intrusion detection system based on a machine learning method.
  • the anti-sample generator should be a black box attacker, and cannot directly obtain the same data as the industrial control intrusion detection system (detector).
  • Different effective features can be extracted for different industrial control system communication protocols.
  • Commonly used industrial control system communication protocols include Modbus, PROFIBUS, DNP3, BACnet, Siemens S7, etc.
  • Each communication protocol has a corresponding format and application scenario, and different communication protocols can be parsed according to the specific scenario to obtain an effective feature set.
  • step (3) Establish a machine learning classifier based on the effective features extracted in step (2), train the machine learning classifier using the labeled industrial control system communication data, and obtain a trained classifier for distinguishing normal or abnormal communication data .
  • the classifier used in the training of the adversarial sample generator can be a classifier different from the industrial control intrusion detection system, that is, the detection party.
  • the classifier generated by the adversarial sample generator can be called a local alternative model of adversarial learning, the principle of which is adversarial learning The portability of the attack.
  • step (3) Using the classifier established in step (3), transform the adversarial learning problem of the industrial control intrusion detection system into the following optimization problem, and solve the optimization problem to obtain the final adversarial sample.
  • g(x) is the probability that the adversarial sample x * is judged to be an abnormal sample, calculated by the classifier
  • d(x * , x 0 ) is the distance between the adversarial sample and the original attack sample
  • d max indicates that the industrial control system allows The maximum Euclidean distance is that exceeding this distance has no malicious effect against the sample.
  • the specific optimization problem solution can use gradient descent method, Newton method, COBYLA (for Constrained Optimization, BY Linear Approximations) method, etc.
  • the expression method of distance can adopt one-paradigm distance, two-paradigm distance, infinite-paradigm distance, etc.
  • the machine learning classifier can use a neural network.
  • the probability calculation method is as follows:
  • step (4) Test the countermeasure sample generated in step (4) in the actual industrial control system. If it can successfully bypass the industrial control intrusion detection system and retain the attack effect, it is regarded as an effective countermeasure sample, otherwise the countermeasure sample is discarded.
  • the initial attack samples include injection attacks, function code attacks, and eavesdropping attacks.
  • the neural network it is designed to solve the problem optimally, and the usage constraints of the specific application scenario are added, such as the fixed selection value of the function code, and the other network features are discrete positive integer values;

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Computer Security & Cryptography (AREA)
  • Computer Hardware Design (AREA)
  • Automation & Control Theory (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Virology (AREA)
  • Quality & Reliability (AREA)
  • Manufacturing & Machinery (AREA)
  • Data Exchanges In Wide-Area Networks (AREA)
  • Computer And Data Communications (AREA)

Abstract

本发明公开了一种基于对抗学习的工业控制系统恶意样本生成方法。该方法通过对抗学习技术,利用优化解法计算出基于机器学习方法的工业控制入侵检测系统的对抗样本。原来能被入侵检测系统发现的攻击样本,经过该方法处理后能生成对应的新对抗样本。这种对抗样本可以绕过原有的入侵检测器(被识别为正常),仍保持攻击效果。本发明通过主动生成工业控制系统恶意样本,有效保障工业控制系统安全,防患于未然。

Description

一种基于对抗学习的工业控制系统恶意样本生成方法 技术领域
本发明属于工业控制系统安全领域,涉及一种基于对抗学习的工业控制系统恶意样本生成方法。
背景技术
频频发生的工业控制系统安全事件引起世界各国的高度重视。为了积极应对工控安全问题,各种应对措施纷纷展开,如规范标准的制定、入侵检测的部署。在工业控制系统中,基于网络的入侵检测器是非常典型且常见的。通过在应用网络(IT)和控制网络(OT)之间部署网络异常检测装置,可以极大地提高工控系统的安全性。目前,基于机器学习的异常检测算法可以在网络异常流量识别中实现较好的性能。这些机器学习技术能够从大量工业控制系统历史数据中找出其内在行为模式以构建异常检测系统。
而另一方面,现有的基于机器学习算法的工业控制系统入侵检测器存在脆弱的地方。机器学习模型在实际应用的时候很容易被攻击者故意生成的对抗样本操纵。攻击者通过对测试示例进行微小的修改,就能让算法模型产生与正常输出完全不同的结果。让模型产生错误预测结果的样本叫做对抗样本。主动对机器学习设计生成对抗性攻击的研究被称为对抗机器学习研究。这种对抗攻击给使用机器学习的系统带来了潜在的安全威胁,尤其是对于诸如工业控制等对安全需求很高的系统。
对于基于机器学习算法的工业控制系统入侵检测系统,如果受到攻击者蓄意的对抗攻击,会无法正常识别出原本能识别的恶意流量。这会给工控系统带来很大的安全隐患。目前现有的工控安全入侵检测研究缺少对于对抗学习部分的研究,而模拟攻击者主动生成针对工控机器学习检测算法的对抗样本是十分有意义的。因此,有必要对工业控制入侵检测系统展开对抗机器学习研究,主动生成能绕过检测器完成隐蔽性攻击的对抗样本,从而保证系统安全,防患于未然。
发明内容
本发明的目的在于针对目前工业控制系统安全的欠缺和不足,提供一种基于对抗学习的工业控制系统恶意样本生成方法。
本发明的目的是通过以下技术方案来实现的:一种基于对抗学习的工业控制系统恶意样本生成方法,包括如下步骤:
(1)对抗样本生成者对工业控制系统通信数据进行嗅探,获得与工业控制入侵检测系统所使用的训练数据同分布的通信数据,并对通信数据标记类别标签,类别包括异常和正常,其中的异常通信数据作为原始攻击样本。工业控制入侵检测系统为现有的基于机器学习方法的工业控制入侵检测系统。
(2)对工业控制系统通信数据进行协议解析工作,识别、提取其中有效特征,包括:通信数据的源IP地址(SIP)、源端口号(SP)、目的IP地址(DIP)、目的端口号(DP)、包间隔时间(Time delta)、包发送时间(Time)、包功能码(Function code)等。
(3)根据步骤(2)提取的有效特征建立机器学习分类器,利用有标签的工业控制系统通信数据对机器学习分类器进行训练,得到训练好的用于分辨正常或异常通信数据的分类器。
(4)利用步骤(3)建立的分类器,将工业控制入侵检测系统的对抗学习问题转化成如下优化问题,求解优化问题得到最终的对抗样本。
x *=arg min g(x)
s.t. d(x *,x 0)<d max
其中g(x)为对抗样本x *被判为异常样本的可能性,通过分类器计算得到;d(x *,x 0)为对抗样本和原始攻击样本的距离,d max表示工业控制系统允许的最大欧式距离,即超过此距离对抗样本无恶意效果。
(5)将步骤(4)生成的对抗样本在实际工业控制系统中进行测试,若能成功绕过工业控制入侵检测系统并且保留攻击效果则视为有效的对抗样本,否则舍弃该对抗样本。
进一步地,所述步骤(1)中,对抗样本生成者应为黑盒攻击者,不能直接获得与工业控制入侵检测系统(检测方)完全相同的数据。
进一步地,所述步骤(2)中,针对不同的工业控制系统通信协议可以提取出不同的有效特征。常用的工业控制系统通信协议包括Modbus、PROFIBUS、DNP3、BACnet、Siemens S7等。其中每种通信协议有对应的格式与应用场景,根据具体场景可以解析不同的通信协议,得出有效特征集合。
进一步地,所述步骤(3)中,对抗样本生成者训练所用的分类器可以是和工业控制入侵检测系统即检测方不一样的分类器,对抗样本生成者生成的分类器可被称为对抗学习的本地替代模型,其原理是对抗学习攻击的可迁移性。
进一步地,所述步骤(4)中,具体优化问题解法包括梯度下降法、牛顿法、COBYLA(for Constrained Optimization BY Linear Approximations)方法等。
进一步地,所述步骤(4)中,距离的表达方法包括一范式距离、二范式距离、无穷范式距离。
进一步地,所述步骤(4)中,机器学习分类器采用神经网络,其概率计算方式如下:
Figure PCTCN2019101247-appb-000001
其中p为预测概率,x (i)为样本x的第i个特征,y为样本x所对应的标签j,θ为神经网络参数,θ j为标签j对应的神经网络参数,k为标签的总数;
将工业控制入侵检测系统的对抗学习问题转化成如下优化问题:
x *=-arg min[p(x)=0]
s.t. d(x *,x 0)<d max
进一步地,所述步骤(4)中,对于特定控制场景,需要在优化问题中加入对变量的特殊约束,如某场景的包功能码是确定的,则对抗样本不能在该维度上做改变。即该方法在应用时,为保证生成的对抗样本能有效的完成恶意攻击效果,生成者需要根据特定的场景对特定维度的变量在做优化问题设计时加上不同的约束条件。
本发明的有益效果是:基于对抗学习的工业控制系统恶意样本生成方法,增加了工业控制系统安全性。将原有的对抗学习研究领域从机器视觉、语音等扩展到了工业控制领域。增强了基于机器学习的工业控制入侵检测系统安全性能,防止恶意样本的攻击。
附图说明
图1是本发明方法实现框图。
图2是分发明方法仿真结果图。
具体实施方式
下面结合附图和具体实施例对本发明作进一步详细说明。
本发明提供的一种基于对抗学习的工业控制系统恶意样本生成方法,如图1所示,包括如下步骤:
(1)对抗样本生成者对工业控制系统通信数据进行嗅探,获得与工业控制入侵检测系统所使用的训练数据同分布的通信数据,并对通信数据标记类别标签,类别包括异常和正常,其中的异常通信数据作为原始攻击样本。工业控制入侵检测系统为现有的基于机器学习方法的工业控制入侵检测系统。
其中,对抗样本生成者应为黑盒攻击者,不能直接获得与工业控制入侵检测系统(检测方)完全相同的数据。
(2)对工业控制系统通信数据进行协议解析工作,识别、提取其中有效特征,包括:通信数据的源IP地址(SIP)、源端口号(SP)、目的IP地址(DIP)、目的端口号(DP)、包间隔时间(Time delta)、包发送时间(Time)、包功能码(Function code)等。
针对不同的工业控制系统通信协议可以提取出不同的有效特征。常用的工业控制系统通信协议包括Modbus、PROFIBUS、DNP3、BACnet、Siemens S7等。其中每种通信协议有对应的格式与应用场景,根据具体场景可以解析不同的通信协议,得出有效特征集合。
(3)根据步骤(2)提取的有效特征建立机器学习分类器,利用有标签的工业控制系统通信数据对机器学习分类器进行训练,得到训练好的用于分辨正常或异常通信数据的分类器。
对抗样本生成者训练所用的分类器可以是和工业控制入侵检测系统即检测方不一样的分类器,对抗样本生成者生成的分类器可被称为对抗学习的本地替代模型,其原理是对抗学习攻击的可迁移性。
(4)利用步骤(3)建立的分类器,将工业控制入侵检测系统的对抗学习问题转化成如下优化问题,求解优化问题得到最终的对抗样本。
x *=arg min g(x)
s.t. d(x *,x 0)<d max
其中g(x)为对抗样本x *被判为异常样本的可能性,通过分类器计算得到;d(x *,x 0)为对抗样本和原始攻击样本的距离,d max表示工业控制系统允许的最大欧式距离,即超过此距离对抗样本无恶意效果。
具体优化问题解法可以采用梯度下降法、牛顿法、COBYLA(for Constrained Optimization BY Linear Approximations)方法等。距离的表达方法可以采用一范式距离、二范式距离、无穷范式距离等。
对于特定控制场景,需要在优化问题中加入对变量的特殊约束,如某场景的包功能码是确定的,则对抗样本不能在该维度上做改变。即该方法在应用时,为保证生成的对抗样本能有效的完成恶意攻击效果,生成者需要根据特定的场景对特定维度的变量在做优化问题设计时加上不同的约束条件。
机器学习分类器可以采用神经网络,采用神经网络时,其概率计算方式如下:
Figure PCTCN2019101247-appb-000002
其中p为预测概率,x (i)为样本x的第i个特征,y为样本x所对应的标签j,θ为神经网络参数,θ j为标签j对应的神经网络参数,k为标签的总数;此时,将工业控制入侵检测系统的对抗学习问题转化成如下优化问题:
x *=-arg min[p(x)=0]
s.t. d(x *,x 0)<d max
(5)将步骤(4)生成的对抗样本在实际工业控制系统中进行测试,若能成功绕过工业控制入侵检测系统并且保留攻击效果则视为有效的对抗样本,否则舍弃该对抗样本。
以下以一个具体的应用场景为例,产生工业控制入侵检测系统对抗样本过程如下:
1.嗅探得到现有的基于机器学习算法的工业控制系统入侵检测器所使用的通信数据;初始攻击样本包括注入攻击、功能码攻击、窃听攻击。
2.进行协议解析,如西门子S7comm协议,解析得到源IP,目的IP,端口号,功能码,子功能码,包间隔时间等特征;
3.在本地生成替代分类器,如用多层感知机生成基本的神经网络算法;
4.根据该神经网络设计成优化求解问题,并加入该具体应用场景的使用约束,如功能码的固定选值,其他网络特征为离散正整数值等;
5.利用COBYLA方法计算得到对抗样本,并在工业控制系统安全测试平台中测试其对抗效果,三种初始攻击样本的攻击成功率见图2。通过图2可以看出,本发明方法对窃听攻击的攻击成功率达到100%,对功能码攻击的攻击成功率达到80%,注入攻击由于实际攻击复杂,较难完成原始攻击样本与对抗样本的转化,其攻击成功率仍然能够达到20%。
上述实施例用来解释说明本发明,而不是对本发明进行限制,在本发明的精神和权利要求的保护范围内,对本发明作出的任何修改和改变,都落入本发明的保护范围。

Claims (8)

  1. 一种基于对抗学习的工业控制系统恶意样本生成方法,其特征在于,包含以下步骤;
    (1)对抗样本生成者对工业控制系统通信数据进行嗅探,获得与工业控制入侵检测系统所使用的训练数据同分布的通信数据,并对通信数据标记类别标签,其中的异常通信数据作为原始攻击样本。
    (2)对工业控制系统通信数据进行协议解析工作,识别、提取其中有效特征,包括:通信数据的源IP地址(SIP)、源端口号(SP)、目的IP地址(DIP)、目的端口号(DP)、包间隔时间(Time delta)、包发送时间(Time)、包功能码(Function code)等。
    (3)根据步骤(2)提取的有效特征建立机器学习分类器,利用有标签的工业控制系统通信数据对机器学习分类器进行训练,得到训练好的用于分辨正常或异常通信数据的分类器。
    (4)利用步骤(3)建立的分类器,将工业控制入侵检测系统的对抗学习问题转化成如下优化问题,求解优化问题得到最终的对抗样本。
    x *=arg min g(x)
    s.t.d(x *,x 0)<d max
    其中g(x)为对抗样本x *被判为异常样本的可能性,通过分类器计算得到;d(x *,x 0)为对抗样本和原始攻击样本的距离,d max表示工业控制系统允许的最大欧式距离,即超过此距离对抗样本无恶意效果。
    (5)将步骤(4)生成的对抗样本在实际工业控制系统中进行测试,若能成功绕过工业控制入侵检测系统并且保留攻击效果则视为有效的对抗样本,否则舍弃该对抗样本。
  2. 根据权利要求1所述的一种基于对抗学习的工业控制系统恶意样本生成方法,其特征在于,所述步骤(1)中,对抗样本生成者应为黑盒攻击者,不能直接获得与工业控制入侵检测系统(检测方)完全相同的数据。
  3. 根据权利要求1所述的一种基于对抗学习的工业控制系统恶意样本生成方法,其特征在于,所述步骤(2)中,针对不同的工业控制系统通信协议可以提取出不同的有效特征。常用的工业控制系统通信协议包括Modbus、PROFIBUS、DNP3、BACnet、Siemens S7等。其中每种通信协议有对应的格式与应用场景,根据具体场景可以解析不同的通信协议,得出有效特征集合。
  4. 根据权利要求1所述的一种基于对抗学习的工业控制系统恶意样本生成方法,其特征在于,所述步骤(3)中,对抗样本生成者训练所用的分类器可以是和工业控制入侵检测系统不 一样的分类器,对抗样本生成者生成的分类器可被称为对抗学习的本地替代模型,其原理是对抗学习攻击的可迁移性。
  5. 根据权利要求1所述的一种基于对抗学习的工业控制系统恶意样本生成方法,其特征在于,所述步骤(4)中,具体优化问题解法包括梯度下降法、牛顿法、COBYLA(for Constrained Optimization BY Linear Approximations)方法等。
  6. 根据权利要求1所述的一种基于对抗学习的工业控制系统恶意样本生成方法,其特征在于,所述步骤(4)中,距离的表达方法包括一范式距离、二范式距离、无穷范式距离。
  7. 根据权利要求1所述的一种基于对抗学习的工业控制系统恶意样本生成方法,其特征在于,所述步骤(4)中,机器学习分类器采用神经网络,其概率计算方式如下:
    Figure PCTCN2019101247-appb-100001
    其中p为预测概率,x (i)为样本x的第i个特征,y为样本x所对应的标签j,θ为神经网络参数,θ j为标签j对应的神经网络参数,k为标签的总数;
    将工业控制入侵检测系统的对抗学习问题转化成如下优化问题:
    x *=-arg min[p(x)=0]
    s.t.d(x *,x 0)<d max
  8. 根据权利要求1所述的一种基于对抗学习的工业控制系统恶意样本生成方法,其特征在于,所述步骤(4)中,对于特定控制场景,需要在优化问题中加入对变量的特殊约束,即该方法在应用时,为保证生成的对抗样本能有效的完成恶意攻击效果,生成者需要根据特定的场景对特定维度的变量在做优化问题设计时加上不同的约束条件。
PCT/CN2019/101247 2019-01-07 2019-08-18 一种基于对抗学习的工业控制系统恶意样本生成方法 Ceased WO2020143227A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
US16/982,056 US20210319113A1 (en) 2019-01-07 2019-08-18 Method for generating malicious samples against industrial control system based on adversarial learning
JP2020540726A JP6932270B2 (ja) 2019-01-07 2019-08-18 敵対的学習に基づく工業制御システムの悪意あるサンプルの生成方法

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910011642.4A CN109902709B (zh) 2019-01-07 2019-01-07 一种基于对抗学习的工业控制系统恶意样本生成方法
CN201910011642.4 2019-01-07

Publications (1)

Publication Number Publication Date
WO2020143227A1 true WO2020143227A1 (zh) 2020-07-16

Family

ID=66943599

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/101247 Ceased WO2020143227A1 (zh) 2019-01-07 2019-08-18 一种基于对抗学习的工业控制系统恶意样本生成方法

Country Status (4)

Country Link
US (1) US20210319113A1 (zh)
JP (1) JP6932270B2 (zh)
CN (1) CN109902709B (zh)
WO (1) WO2020143227A1 (zh)

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112115467A (zh) * 2020-09-04 2020-12-22 长沙理工大学 一种基于集成学习的半监督分类的入侵检测方法
CN112162515A (zh) * 2020-10-10 2021-01-01 浙江大学 一种针对过程监控系统的对抗攻击方法
CN112381152A (zh) * 2020-11-17 2021-02-19 上海科技大学 一种基于攻击成本的对抗样本检测方法
CN112465019A (zh) * 2020-11-26 2021-03-09 重庆邮电大学 一种基于扰动的对抗样本生成与对抗性防御方法
CN113221100A (zh) * 2021-02-09 2021-08-06 上海大学 一种面向工业互联网边界防护的对抗式入侵检测方法
CN114492537A (zh) * 2022-02-15 2022-05-13 中国科学技术大学 一种肌电控制系统遭恶意攻击的异常信号检测方法和相关设备
CN114895559A (zh) * 2022-04-18 2022-08-12 北京科技大学 一种面向恶意攻击下的信息物理系统可靠控制方法
CN115240660A (zh) * 2022-05-31 2022-10-25 宁波大学 一种基于帧偏移的语音对抗样本防御方法
CN115620160A (zh) * 2022-10-21 2023-01-17 中国地质大学(武汉) 基于多分类器对抗主动迁移学习的遥感图像分类方法
WO2023286269A1 (ja) * 2021-07-16 2023-01-19 日本電気株式会社 学習用データ生成装置、学習用データの生成方法、プログラム、検知モデルの生成方法および認証システム
CN116226851A (zh) * 2023-02-14 2023-06-06 湖南大学 推荐系统攻击检测样本数据生成方法与装置
CN116318877A (zh) * 2023-02-16 2023-06-23 北京理工大学 利用多种特征流形向量的入侵检测系统对抗样本防御方法
CN117172304A (zh) * 2023-09-21 2023-12-05 湘潭大学 一种高迁移性和弱感知性的对抗样本生成方法

Families Citing this family (35)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109902709B (zh) * 2019-01-07 2020-12-08 浙江大学 一种基于对抗学习的工业控制系统恶意样本生成方法
CN110633570B (zh) * 2019-07-24 2021-05-11 浙江工业大学 面向恶意软件汇编格式检测模型的黑盒攻击的防御方法
CN112311733A (zh) * 2019-07-30 2021-02-02 四川大学 一种基于强化学习优化xss检测模型防御对抗攻击的方法
CN110619216B (zh) * 2019-09-17 2021-09-03 武汉思普崚技术有限公司 一种对抗性网络的恶意软件检测方法及系统
CN111104982B (zh) * 2019-12-20 2021-09-24 电子科技大学 一种标签无关的跨任务对抗样本生成方法
US11811791B2 (en) * 2020-01-09 2023-11-07 Vmware, Inc. Generative adversarial network based predictive model for collaborative intrusion detection systems
CN111242322B (zh) * 2020-04-24 2020-08-14 支付宝(杭州)信息技术有限公司 后门样本的检测方法、装置和电子设备
US20220038474A1 (en) * 2020-07-31 2022-02-03 Dev Ayan Nag Method and system that enhances computer-system security by identifying and blocking harmful communications through component interfaces
CN112333402B (zh) 2020-10-20 2021-10-22 浙江大学 一种基于声波的图像对抗样本生成方法及系统
US11818147B2 (en) * 2020-11-23 2023-11-14 Fair Isaac Corporation Overly optimistic data patterns and learned adversarial latent features
JP7740812B2 (ja) * 2021-02-25 2025-09-17 ロバスト インテリジェンス インコーポレイテッド 人工知能モデルを安全に配備するための方法及びシステム
CN115225295A (zh) * 2021-04-16 2022-10-21 华为技术有限公司 攻击样本管理的方法以及设备
CN112819109B (zh) * 2021-04-19 2021-06-18 中国工程物理研究院计算机应用研究所 针对黑盒对抗样本攻击的视频分类系统安全性增强方法
US20220414531A1 (en) * 2021-06-25 2022-12-29 International Business Machines Corporation Mitigating adversarial attacks for simultaneous prediction and optimization of models
CN113938309A (zh) * 2021-10-28 2022-01-14 西安热工研究院有限公司 一种基于DCGAN网络的DoS攻击流量生成系统及方法
CN114139601B (zh) * 2021-11-01 2025-03-07 国家电网有限公司大数据中心 一种对电力巡检场景人工智能算法模型的评估方法及系统
CN114254694B (zh) * 2021-11-15 2025-04-04 中国科学院信息工程研究所 针对新型天基网络中入侵检测系统对抗训练的方法及装置
CN114444075B (zh) * 2022-02-09 2023-05-19 深圳市前海新型互联网交换中心有限公司 一种生成躲避流量数据的方法
CN114861762B (zh) * 2022-04-08 2026-04-17 联合汽车电子有限公司 一种信息处理方法、装置、存储介质及车载电子控制器
CN115277065B (zh) * 2022-06-15 2024-01-23 北京信息科技大学 一种物联网异常流量检测中的对抗攻击方法及装置
CN115499222B (zh) * 2022-09-20 2024-11-22 中国农业银行股份有限公司 恶意请求检测模型的训练方法、恶意请求识别方法及设备
CN115909020B (zh) * 2022-09-30 2024-01-09 北京瑞莱智慧科技有限公司 模型鲁棒性检测方法、相关装置及存储介质
CN115659334B (zh) * 2022-10-17 2025-07-15 华中科技大学 一种智能合约对抗样本生成方法
CN115333874B (zh) * 2022-10-18 2023-04-28 北京珞安科技有限责任公司 一种工业终端主机监测方法
CN115766193B (zh) * 2022-11-11 2025-08-19 浙江大学 基于无监督学习的工业网络异常流量深度检测方法
CN116527373B (zh) * 2023-05-18 2023-10-20 清华大学 针对恶意url检测系统的后门攻击方法和装置
CN116304959B (zh) * 2023-05-24 2023-08-15 山东省计算中心(国家超级计算济南中心) 一种用于工业控制系统的对抗样本攻击防御方法及系统
CN116668112B (zh) * 2023-05-29 2025-10-17 广州大学 一种生成流量对抗样本访问黑盒模型的方法及装置
CN116996266B (zh) * 2023-07-05 2026-05-19 中国电信股份有限公司技术创新中心 一种apt攻击数据的生成方法及相关装置
CN118337526B (zh) * 2024-06-11 2024-09-13 长春大学 一种对抗攻击样本生成方法
CN118485114B (zh) * 2024-07-15 2024-09-27 湖南大学 一种基于元学习的生成式对抗攻击方法
CN119603026B (zh) * 2024-11-28 2025-07-11 西南交通大学 基于入侵攻击对抗样本的网络安全防御方法
CN119363857B (zh) * 2024-12-20 2025-04-11 中汽数据(天津)有限公司 一种汽车数据传输监测方法和系统
CN120395818A (zh) * 2025-04-18 2025-08-01 北京科技大学 一种基于生成对抗模仿学习的机械臂轨迹规划方法及装置
CN120692080A (zh) * 2025-06-26 2025-09-23 北京计算机技术及应用研究所 一种基于对抗性机器学习的网络流量审计优化防御方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2006007633A1 (en) * 2004-07-16 2006-01-26 National Ict Australia Limited Data mining unlearnable data sets
CN108171266A (zh) * 2017-12-25 2018-06-15 中国矿业大学 一种多目标深度卷积生成式对抗网络模型的学习方法
CN108509966A (zh) * 2017-02-27 2018-09-07 顾泽苍 一种超深度对抗学习的方法
CN109902709A (zh) * 2019-01-07 2019-06-18 浙江大学 一种基于对抗学习的工业控制系统恶意样本生成方法

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10043261B2 (en) * 2016-01-11 2018-08-07 Kla-Tencor Corp. Generating simulated output for a specimen
US20190294803A1 (en) * 2016-12-01 2019-09-26 Mitsubishi Electric Corporation Evaluation device, security product evaluation method, and computer readable medium
KR102403494B1 (ko) * 2017-04-27 2022-05-27 에스케이텔레콤 주식회사 생성적 대립 네트워크에 기반한 도메인 간 관계를 학습하는 방법
CN107819790A (zh) * 2017-12-08 2018-03-20 中盈优创资讯科技有限公司 攻击报文的识别方法及装置
CN108322349B (zh) * 2018-02-11 2021-04-06 浙江工业大学 基于对抗式生成网络的深度学习对抗性攻击防御方法
CN108537271B (zh) * 2018-04-04 2021-02-05 重庆大学 一种基于卷积去噪自编码机防御对抗样本攻击的方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2006007633A1 (en) * 2004-07-16 2006-01-26 National Ict Australia Limited Data mining unlearnable data sets
CN108509966A (zh) * 2017-02-27 2018-09-07 顾泽苍 一种超深度对抗学习的方法
CN108171266A (zh) * 2017-12-25 2018-06-15 中国矿业大学 一种多目标深度卷积生成式对抗网络模型的学习方法
CN109902709A (zh) * 2019-01-07 2019-06-18 浙江大学 一种基于对抗学习的工业控制系统恶意样本生成方法

Cited By (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112115467A (zh) * 2020-09-04 2020-12-22 长沙理工大学 一种基于集成学习的半监督分类的入侵检测方法
CN112162515A (zh) * 2020-10-10 2021-01-01 浙江大学 一种针对过程监控系统的对抗攻击方法
CN112162515B (zh) * 2020-10-10 2021-08-03 浙江大学 一种针对过程监控系统的对抗攻击方法
CN112381152A (zh) * 2020-11-17 2021-02-19 上海科技大学 一种基于攻击成本的对抗样本检测方法
CN112381152B (zh) * 2020-11-17 2023-12-05 上海科技大学 一种基于攻击成本的对抗样本检测方法
CN112465019A (zh) * 2020-11-26 2021-03-09 重庆邮电大学 一种基于扰动的对抗样本生成与对抗性防御方法
CN113221100A (zh) * 2021-02-09 2021-08-06 上海大学 一种面向工业互联网边界防护的对抗式入侵检测方法
WO2023286269A1 (ja) * 2021-07-16 2023-01-19 日本電気株式会社 学習用データ生成装置、学習用データの生成方法、プログラム、検知モデルの生成方法および認証システム
CN114492537A (zh) * 2022-02-15 2022-05-13 中国科学技术大学 一种肌电控制系统遭恶意攻击的异常信号检测方法和相关设备
CN114895559A (zh) * 2022-04-18 2022-08-12 北京科技大学 一种面向恶意攻击下的信息物理系统可靠控制方法
CN115240660A (zh) * 2022-05-31 2022-10-25 宁波大学 一种基于帧偏移的语音对抗样本防御方法
CN115620160A (zh) * 2022-10-21 2023-01-17 中国地质大学(武汉) 基于多分类器对抗主动迁移学习的遥感图像分类方法
CN115620160B (zh) * 2022-10-21 2025-12-02 中国地质大学(武汉) 基于多分类器对抗主动迁移学习的遥感图像分类方法
CN116226851A (zh) * 2023-02-14 2023-06-06 湖南大学 推荐系统攻击检测样本数据生成方法与装置
CN116318877A (zh) * 2023-02-16 2023-06-23 北京理工大学 利用多种特征流形向量的入侵检测系统对抗样本防御方法
CN117172304A (zh) * 2023-09-21 2023-12-05 湘潭大学 一种高迁移性和弱感知性的对抗样本生成方法

Also Published As

Publication number Publication date
CN109902709B (zh) 2020-12-08
JP6932270B2 (ja) 2021-09-08
JP2021513143A (ja) 2021-05-20
CN109902709A (zh) 2019-06-18
US20210319113A1 (en) 2021-10-14

Similar Documents

Publication Publication Date Title
CN109902709B (zh) 一种基于对抗学习的工业控制系统恶意样本生成方法
CN111181901B (zh) 异常流量检测装置及其异常流量检测方法
CN109768952B (zh) 一种基于可信模型的工控网络异常行为检测方法
CN113268735B (zh) 分布式拒绝服务攻击检测方法、装置、设备和存储介质
CN107360145B (zh) 一种多节点蜜罐系统及其数据分析方法
Lee et al. Detection of DDoS attacks using optimized traffic matrix
Garitano et al. A review of SCADA anomaly detection systems
Li et al. Prism: Real-time privacy protection against temporal network traffic analyzers
CN105871861B (zh) 一种自学习协议规则的入侵检测方法
CN112149120A (zh) 一种透传式双通道电力物联网安全检测系统
CN111669354A (zh) 基于机器学习的威胁情报工业防火墙
CN121239440A (zh) 基于深度学习的sip协议加密恶意流量检测方法
Thiyagarajan et al. Anomaly detection in air-gapped industrial control systems of nuclear power plants
Johnstone et al. Timing attack detection on BACnet via a machine learning approach
CN108494791A (zh) 一种基于Netflow日志数据的DDOS攻击检测方法及装置
CN117220953A (zh) 面向网络入侵检测系统的对抗性流量生成方法及系统
CN108040052A (zh) 一种基于Netflow日志数据的网络安全威胁分析方法及系统
CN116506216B (zh) 一种轻量化恶意流量检测存证方法、装置、设备及介质
Oachesu Enhancing Smart Home Security with Generative AI: Threat Detection and Countermeasure Design
CN119996082B (zh) 一种基于大语言模型来实现网络安全威胁预测的方法
Yu et al. Mining anomaly communication patterns for industrial control systems
Wang et al. Intrusion detection model of SCADA using graphical features
Tolkachov et al. DEVELOPMENT OF A MODEL FOR THE ANALYSIS AND SEPARATION OF SERVICE AND USEFUL TRAFFIC IN CYBER-PHYSICAL SYSTEMS.
Mykhailenko et al. ENHANCING CYBERSECURITY WITH AI-DRIVEN SIEM SYSTEMS
Yu et al. A Hierarchical Intrusion Detection System for Industrial Control Networks based on EtherNet/IP

Legal Events

Date Code Title Description
ENP Entry into the national phase

Ref document number: 2020540726

Country of ref document: JP

Kind code of ref document: A

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19908432

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19908432

Country of ref document: EP

Kind code of ref document: A1