Biblio

In control systems, the operation of the system after an incident occurs is important. This paper proposes to design a whitelist model that can detect anomalies and identify locations of anomalous actuators using finite automata during multiple actuators attack. By applying this model and comparing the whitelist model with the operation data, the monitoring system detects anomalies and identifies anomaly locations of actuator that deviate from normal operation. We propose to construct a whitelist model focusing on the order of the control system operation using binary search trees, which can grasp the state of the system when anomalies occur. We also apply combinatorial compression based on BDD (Binary Decision Diagram) to the model to speed up querying and identification of abnormalities. Based on the model designed in this study, we aim to construct a secured control system that selects and executes an appropriate fallback operation based on the state of the system when anomaly is detected.

Recent approaches have proven the effectiveness of local outlier factor-based outlier detection when applied over traffic flow probability distributions. However, these approaches used distance metrics based on the Bhattacharyya coefficient when calculating probability distribution similarity. Consequently, the limited expressiveness of the Bhattacharyya coefficient restricted the accuracy of the methods. The crucial deficiency of the Bhattacharyya distance metric is its inability to compare distributions with non-overlapping sample spaces over the domain of natural numbers. Traffic flow intensity varies greatly, which results in numerous non-overlapping sample spaces, rendering metrics based on the Bhattacharyya coefficient inappropriate. In this work, we address this issue by exploring alternative distance metrics and showing their applicability in a massive real-life traffic flow data set from 26 vital intersections in The Hague. The results on these data collected from 272 sensors for more than two years show various advantages of the Earth Mover's distance both in effectiveness and efficiency.

We propose Cleann, the first end-to-end framework that enables online mitigation of Trojans for embedded Deep Neural Network (DNN) applications. A Trojan attack works by injecting a backdoor in the DNN while training; during inference, the Trojan can be activated by the specific backdoor trigger. What differentiates Cleann from the prior work is its lightweight methodology which recovers the ground-truth class of Trojan samples without the need for labeled data, model retraining, or prior assumptions on the trigger or the attack. We leverage dictionary learning and sparse approximation to characterize the statistical behavior of benign data and identify Trojan triggers. Cleann is devised based on algorithm/hardware co-design and is equipped with specialized hardware to enable efficient real-time execution on resource-constrained embedded platforms. Proof of concept evaluations on Cleann for the state-of-the-art Neural Trojan attacks on visual benchmarks demonstrate its competitive advantage in terms of attack resiliency and execution overhead.

In the industrial control system, in order to solve the problem that the installation of smart devices in a transparent environment are faced with the unknown attack problems, because most of the industrial control equipment to detect unknown risks, Therefore, by studying the security protection of the current industrial control system and the trust mechanism that should be widely used in the Internet of things, this paper presents the abnormal node detection mode based on comprehensive trust applied to the industrial control system scenarios. This model firstly proposes a model, which fuses direct and indirect trust values into current trust values through support algorithm and vector similarity algorithm, and then combines them with historical trust values, and gives the calculation method of each trust value. Finally, a method to determine abnormal nodes based on comprehensive trust degree is given to realize a detection process for all industrial control nodes. By analyzing the real data case provided by Melbourne Water, it is concluded that this model can improve the detection range and detection accuracy of abnormal nodes. It can accurately judge and effectively resist malicious behavior and also have a good resistance to aggression.

Industrial Control System (ICS) transmits control and monitoring data between devices in an industrial environment that includes smart grids, water and gas distribution, or traffic control. Unlike traditional internet communication, ICS traffic is stable, periodical, and with regular communication patterns that can be described using statistical modeling. By observing selected features of ICS transmission, e.g., packet direction and inter-arrival times, we can create a statistical profile of the communication based on distribution of features learned from the normal ICS traffic. This paper demonstrates that using statistical modeling, we can detect various anomalies caused by irregular transmissions, device or link failures, and also cyber attacks like packet injection, scanning, or denial of service (DoS). The paper shows how a statistical model is automatically created from a training dataset. We present two types of statistical profiles: the master-oriented profile for one-to-many communication and the peer-to-peer profile that describes traffic between two ICS devices. The proposed approach is fast and easy to implement as a part of an intrusion detection system (IDS) or an anomaly detection (AD) module. The proof-of-concept is demonstrated on two industrial protocols: IEC 60870-5-104 (aka IEC 104) and IEC 61850 (Goose).

Industrial Control Systems (ICSs) are used in several domains such as Transportation, Manufacturing, Defense and Power Generation and Distribution. ICSs deal with complex physical systems in order to achieve an industrial purpose with operational safety. Security has not been taken into account by design in these systems that makes them vulnerable to cyberattacks.In this paper, we rely on existing public ICS datasets as well as on the existing literature of Machine Learning (ML) applications for anomaly detection in ICSs in order to improve detection scores. To perform this purpose, we propose a systematic framework, relying on established ML algorithms and suitable data preprocessing methods, which allows us to quickly get efficient, and surprisingly, better results than the literature. Finally, some recommendations for future public ICS dataset generations end this paper, which would be fruitful for improving future attack detection models and then protect new ICSs designed in the next future.

Industrial Control Systems (ICS) are complex systems made up of many components with different tasks. For a safe and secure operation, each device needs to carry out its tasks correctly. To monitor a system and ensure the correct behavior of systems, anomaly detection is used.Models of expected behavior often rely only on cyber or physical features for anomaly detection. We propose an anomaly detection system that combines both types of features to create a dynamic fingerprint of an ICS. We present how a cyber-physical anomaly detection using sound on the physical layer can be designed, and which challenges need to be overcome for a successful implementation. We perform an initial evaluation for identifying actions of a 3D printer.

Industrial Control System (ICS) communication transmits monitoring and control data between industrial processes and the control station. ICS systems cover various domains of critical infrastructure such as the power plants, water and gas distribution, or aerospace traffic control. Security of ICS systems is usually implemented on the perimeter of the network using ICS enabled firewalls or Intrusion Detection Systems (IDSs). These techniques are helpful against external attacks, however, they are not able to effectively detect internal threats originating from a compromised device with malicious software. In order to mitigate or eliminate internal threats against the ICS system, we need to monitor ICS traffic and detect suspicious data transmissions that differ from common operational communication. In our research, we obtain ICS monitoring data using standardized IPFIX flows extended with meta data extracted from ICS protocol headers. Unlike other anomaly detection approaches, we focus on modelling the semantics of ICS communication obtained from the IPFIX flows that describes typical conversational patterns. This paper presents a technique for modelling ICS conversations using frequency prefix trees and Deterministic Probabilistic Automata (DPA). As demonstrated on the attack scenarios, these models are efficient to detect common cyber attacks like the command injection, packet manipulation, network scanning, or lost connection. An important advantage of our approach is that the proposed technique can be easily integrated into common security information and event management (SIEM) systems with Netflow/IPFIX support. Our experiments are performed on IEC 60870-5-104 (aka IEC 104) control communication that is widely used for the substation control in smart grids.

System logs record valuable information about the runtime status of IT systems. Therefore, system logs are a naturally excellent source of information for anomaly detection. Most of the existing studies on log-based anomaly detection construct a detection model to identify anomalous logs. Generally, the model treats historical logs as natural language sequences and learns the normal patterns from normal log sequences, and detects deviations from normal patterns as anomalies. However, the majority of existing methods focus on sequential and quantitative information and ignore semantic information hidden in log sequence so that they are inefficient in anomaly detection. In this paper, we propose a novel framework for automatically detecting log anomalies by utilizing an attention-based Bi-LSTM model. To demonstrate the effectiveness of our proposed model, we evaluate the performance on a public production log dataset. Extensive experimental results show that the proposed approach outperforms all comparison methods for anomaly detection.

With the rapid advancement of the electrical grid, substation automation systems (SASs) have been developing continuously. However, with the introduction of advanced features, such as remote control, potential cyber security threats in SASs are also increased. Additionally, crucial components in SASs, such as protection relays, usually come from third-party vendors and may not be fully trusted. Untrusted devices may stealthily perform harmful or unauthorised behaviours which could compromise or damage SASs, and therefore, bring adverse impacts to the primary plant. Thus, it is necessary to detect abnormal behaviours from an untrusted device before it brings about catastrophic impacts. Anomaly detection techniques are suitable to detect anomalies in SASs as they only bring minimal side-effects to normal system operations. Many researchers have developed various machine learning algorithms and mathematical models to improve the accuracy of anomaly detection. However, without prudent feature selection, it is difficult to achieve high accuracy when detecting attacks launched from internal trusted networks, especially for stealthy message modification attacks which only modify message payloads slightly and imitate patterns of benign behaviours. Therefore, this paper presents choices of features which improve the accuracy of anomaly detection within SASs, especially for detecting “stealthy” attacks. By including two additional features, Boolean control data from message payloads and physical values from sensors, our method improved the accuracy of anomaly detection by decreasing the false-negative rate from 25% to 5% approximately.

With the rapid development of communication net-work, the types and quantities of network traffic data have in-creased substantially. What followed was the frequent occurrence of versatile cyber attacks. As an important part of network security, the network-based intrusion detection system (NIDS) can monitor and protect the network equippments and terminals in real time. The traditional detection methods based on deep learning (DL) are always in supervised manners in NIDS, which can automatically build end-to-end detection model without man-ual feature extraction and selection by domain experts. However, supervised learning methods require large-scale labeled data, yet capturing large labeled datasets is a very cubersome, tedious and time-consuming manual task. Instead, unsupervised learning is an effective way to overcome this problem. Nonetheless, the ex-isting unsupervised methods are prone to low detection efficiency and are difficult to train. In this paper we propose a novel NIDS method called PGAN based on generative adversarial network (GAN) to detect the abnormal traffic from the perspective of Anomaly Detection, which leverage the competitive speciality of adversarial training to learn the normal traffic. Based on the public dataset CICIDS2017, three experimental results show that PGAN can significantly outperform other unsupervised methods like stacked autoencoder (SAE) and isolation forest (IF).

Modern cars are becoming more accessible targets for cyberattacks due to the proliferation of wireless communication channels. The intra-vehicle Controller Area Network (CAN) bus lacks authentication, which exposes critical components to interference from less secure, wirelessly compromised modules. To address this issue, we propose vProfile, a sender authentication system based on voltage fingerprints of Electronic Control Units (ECUs). vProfile exploits the physical properties of ECU output voltages on the CAN bus to determine the authenticity of bus messages, which enables the detection of both hijacked ECUs and external devices connected to the bus. We show the potential of vProfile using experiments on two production vehicles with precision and recall scores of over 99.99%. The improved identification rates and more straightforward design of vProfile make it an attractive improvement over existing methods.

In the ever evolving Internet threat landscape, Distributed Denial-of-Service (DDoS) attacks remain a popular means to invoke service disruption. DDoS attacks, however, have evolved to become a tool of deceit, providing a smokescreen or distraction while some other underlying attack takes place, such as data exfiltration. Knowing the intent of a DDoS, and detecting underlying attacks which may be present concurrently with it, is a challenging problem. An entity whose network is under a DDoS attack may not have the support personnel to both actively fight a DDoS and try to mitigate underlying attacks. Therefore, any system that can detect such underlying attacks should do so only with a high degree of confidence. Previous work utilizing flow aggregation techniques with multi-class anomaly detection showed promise in both DDoS detection and detecting underlying attacks ongoing during an active DDoS attack. In this work, we head in the opposite direction, utilizing flow segmentation and concurrent flow feature aggregation, with the primary goal of greatly reduced detection times of both DDoS and underlying attacks. Using the same multi-class anomaly detection approach, we show greatly improved detection times with promising detection performance.

In this conceptual work, we present Deep Convolutional Gaussian Mixture Models (DCGMMs): a new formulation of deep hierarchical Gaussian Mixture Models (GMMs) that is particularly suitable for describing and generating images. Vanilla (i.e., flat) GMMs require a very large number of components to describe images well, leading to long training times and memory issues. DCGMMs avoid this by a stacked architecture of multiple GMM layers, linked by convolution and pooling operations. This allows to exploit the compositionality of images in a similar way as deep CNNs do. DCGMMs can be trained end-to-end by Stochastic Gradient Descent. This sets them apart from vanilla GMMs which are trained by Expectation-Maximization, requiring a prior k-means initialization which is infeasible in a layered structure. For generating sharp images with DCGMMs, we introduce a new gradient-based technique for sampling through non-invertible operations like convolution and pooling. Based on the MNIST and FashionMNIST datasets, we validate the DCGMMs model by demonstrating its superiority over flat GMMs for clustering, sampling and outlier detection.

In the era of rapid technological growth, malicious traffic has drawn increased attention. Most well-known offensive security assessment todays are heavily focused on pre-compromise. The amount of anomalous data in today's context is massive. Analyzing the data using primitive methods would be highly challenging. Solution to it is: If we can detect adversary behaviors in the early stage of compromise, one can prevent and safeguard themselves from various attacks including ransomwares and Zero-day attacks. Integration of new technologies Artificial Intelligence & Machine Learning with manual Anomaly Detection can provide automated machine-based detection which in return can provide the fast, error free, simplify & scalable Threat Detection & Response System. Endpoint Detection & Response (EDR) tools provide a unified view of complex intrusions using known adversarial behaviors to identify intrusion events. We have used the EMBER dataset, which is a labelled benchmark dataset. It is used to train machine learning models to detect malicious portable executable files. This dataset consists of features derived from 1.1 million binary files: 900,000 training samples among which 300,000 were malicious, 300,000 were benevolent, 300,000 un-labelled, and 200,000 evaluation samples among which 100K were malicious, 100K were benign. We have also included open-source code for extracting features from additional binaries, enabling the addition of additional sample features to the dataset.

Bitcoin is a representative cryptocurrency system using a permissionless peer-to-peer (P2P) network as its communication infrastructure. A number of attacks against Bitcoin have been discovered over the past years, including the Eclipse and EREBUS Attacks. In this paper, we present a new attack against Bitcoin’s P2P networking, dubbed ConMan because it leverages connection manipulation. ConMan achieves the same effect as the Eclipse and EREBUS Attacks in isolating a target (i.e., victim) node from the rest of the Bitcoin network. However, ConMan is different from these attacks because it is an active and deterministic attack, and is more effective and efficient. We validate ConMan through proof-of-concept exploitation in an environment that is coupled with real-world Bitcoin node functions. Experimental results show that ConMan only needs a few minutes to fully control the peer connections of a target node, which is in sharp contrast to the tens of days that are needed by the Eclipse and EREBUS Attacks. Further, we propose several countermeasures against ConMan. Some of them would be effective but incompatible with the design principles of Bitcoin, while the anomaly detection approach is positively achievable. We disclosed ConMan to the Bitcoin Core team and received their feedback, which confirms ConMan and the proposed countermeasures.

Bitcoin has received a lot of attention from investors, researchers, regulators, and the media. It is a known fact that the Bitcoin price usually fluctuates greatly. However, not enough scientific research has been done on these fluctuations. In this study, long short-term memory (LSTM) modeling from Recurrent Neural Networks, which is one of the deep learning methods, was applied on Bitcoin values. As a result of this application, anomaly detection was carried out in the values from the data set. With the LSTM network, a time-dependent representation of Bitcoin price can be captured, and anomalies can be selected. The factors that play a role in the formation of the model to be applied in the detection of anomalies with the experimental results were evaluated.

Data is one of the most valuable assets of an organization and has a tremendous impact on its long-term success and decision-making processes. Typically, organizational data error and outlier detection processes perform manually and reactively, making them time-consuming and prone to human errors. Additionally, rich data types, unlabeled data, and increased volume have made such data more complex. Accordingly, an automated anomaly detection approach is required to improve data management and quality control processes. This study introduces an unsupervised anomaly detection approach based on models comparison, consensus learning, and a combination of rules of thumb with iterative hyper-parameter tuning to increase data quality. Furthermore, a domain expert is considered a human in the loop to evaluate and check the data quality and to judge the output of the unsupervised model. An experiment has been conducted to assess the proposed approach in the context of a case study. The experiment results confirm that the proposed approach can improve the quality of organizational data and facilitate anomaly detection processes.

Insider threat makes enterprises or organizations suffer from the loss of property and the negative influence of reputation. User behavior analysis is the mainstream method of insider threat detection, but due to the lack of fine-grained detection and the inability to effectively capture the behavior patterns of individual users, the accuracy and precision of detection are insufficient. To solve this problem, this paper designs an insider threat detection method based on user historical behavior and attention mechanism, including using Long Short Term Memory (LSTM) to extract user behavior sequence information, using Attention-based on user history behavior (ABUHB) learns the differences between different user behaviors, uses Bidirectional-LSTM (Bi-LSTM) to learn the evolution of different user behavior patterns, and finally realizes fine-grained user abnormal behavior detection. To evaluate the effectiveness of this method, experiments are conducted on the CMU-CERT Insider Threat Dataset. The experimental results show that the effectiveness of this method is 3.1% to 6.3% higher than that of other comparative model methods, and it can detect insider threats in different user behaviors with fine granularity.

Internal attacks are one of the biggest cybersecurity issues to companies and businesses. Despite the implemented perimeter security systems, the risk of adversely affecting the security and privacy of the organization’s information remains very high. Actually, the detection of such a threat is known to be a very complicated problem, presenting many challenges to the research community. In this paper, we investigate the effectiveness and usefulness of using Autoencoder and Variational Autoencoder deep learning algorithms to automatically defend against insider threats, without human intervention. The performance evaluation of the proposed models is done on the public CERT dataset (CERT r4.2) that contains both benign and malicious activities generated from 1000 simulated users. The comparison results with other models show that the Variational Autoencoder neural network provides the best overall performance with a higher detection accuracy and a reasonable false positive rate.

Insider threats are the cyber attacks from the trusted entities within an organization. An insider attack is hard to detect as it may not leave a footprint and potentially cause huge damage to organizations. Anomaly detection is the most common approach for insider threat detection. Lack of real-world data and the skewed class distribution in the datasets makes insider threat analysis an understudied research area. In this paper, we propose a Conditional Generative Adversarial Network (CGAN) to enrich under-represented minority class samples to provide meaningful and diverse data for anomaly detection from the original malicious scenarios. Comprehensive experiments performed on benchmark dataset demonstrates the effectiveness of using CGAN augmented data, and the capability of multi-class anomaly detection for insider activity analysis. Moreover, the method is compared with other existing methods against different parameters and performance metrics.

In recent years, insider threat incidents and losses of companies or organizations are on the rise, and internal network security is facing great challenges. Traditional intrusion detection methods cannot identify malicious behaviors of insiders. As an effective method, insider threat detection technology has been widely concerned and studied. In this paper, we use the tree structure method to analyze user behavior, form feature sequences, and combine the Copula Based Outlier Detection (COPOD) method to detect the difference between feature sequences and identify abnormal users. We experimented on the insider threat dataset CERT-IT and compared it with common methods such as Isolation Forest.

The concealment and confusion nature of insider threat makes it a challenging task for security analysts to identify insider threat from log data. To detect insider threat, we propose a novel gated recurrent unit (GRU) and multi-autoencoder based insider threat detection method, which is an unsupervised anomaly detection method. It takes advantage of the extremely unbalanced characteristic of insider threat data and constructs a normal behavior autoencoder with low reconfiguration error through multi-level filter behavior learning, and identifies the behavior data with high reconfiguration error as abnormal behavior. In order to achieve the high efficiency of calculation and detection, GRU and multi-head attention are introduced into the autoencoder. Use dataset v6.2 of the CERT insider threat as validation data and threat detection recall as evaluation metric. The experimental results show that the effect of the proposed method is obviously better than that of Isolation Forest, LSTM autoencoder and multi-channel autoencoders based insider threat detection methods, and it's an effective insider threat detection technology.

This research concerns the detection of unauthorised access within hospital networks through the real-time analysis of audit logs. Privacy is a primary concern amongst patients due to the rising adoption of Electronic Patient Record (EPR) systems. There is growing evidence to suggest that patients may withhold information from healthcare providers due to lack of Trust in the security of EPRs. Yet, patient record data must be available to healthcare providers at the point of care. Ensuring privacy and confidentiality of that data is challenging. Roles within healthcare organisations are dynamic and relying on access control is not sufficient. Through proactive monitoring of audit logs, unauthorised accesses can be detected and presented to an analyst for review. Advanced data analytics and visualisation techniques can be used to aid the analysis of big data within EPR audit logs to identify and highlight pertinent data points. Employing a human-in-the-loop model ensures that suspicious activity is appropriately investigated and the data analytics is continuously improving. This paper presents a system that employs a Human-in-the-Loop Machine Learning (HILML) algorithm, in addition to a density-based local outlier detection model. The system is able to detect 145 anomalous behaviours in an unlabelled dataset of 1,007,727 audit logs. This equates to 0.014% of the EPR accesses being labelled as anomalous in a specialist Liverpool (UK) hospital.

Static information flow analysis (IFA) and dynamic information flow tracking (DIFT) have been widely employed in offline security analysis of computer programs. As security attacks become more sophisticated, there is a rising need for IFA and DIFT in production environment. However, existing systems usually deal with IFA and DIFT separately, and most DIFT systems incur significant performance overhead. We propose MIT to facilitate IFA and DIFT in online production environment. MIT offers mixed-mode information flow tracking at byte-granularity and incurs moderate runtime performance overhead. The core techniques consist of the extraction of taint semantics intermediate representation (TSIR) at compile-time and the decoupled execution of TSIR for information flow analysis. We conducted an extensive performance overhead evaluation on MIT to confirm its applicability in production environment. We also outline potential applications of MIT, including the implementation of data provenance checking and information flow based anomaly detection in real-world applications.