Biblio

Explainable artificial intelligence (XAI) methods rely on access to model architecture and parameters that is not always feasible for most users, practitioners, and regulators. Inspired by cognitive psychology, we present a case for response times (RTs) as a technique for XAI. RTs are observable without access to the model. Moreover, dynamic inference models performing conditional computation generate variable RTs for visual learning tasks depending on hierarchical representations. We show that MSDNet, a conditional computation model with early-exit architecture, exhibits slower RT for images with more complex features in the ObjectNet test set, as well as the human phenomenon of scene grammar, where object recognition depends on intrascene object-object relationships. These results cast light on MSDNet's feature space without opening the black box and illustrate the promise of RT methods for XAI.

Research into the explanation of machine learning models, i.e., explainable AI (XAI), has seen a commensurate exponential growth alongside deep artificial neural networks throughout the past decade. For historical reasons, explanation and trust have been intertwined. However, the focus on trust is too narrow, and has led the research community astray from tried and true empirical methods that produced more defensible scientific knowledge about people and explanations. To address this, we contribute a practical path forward for researchers in the XAI field. We recommend researchers focus on the utility of machine learning explanations instead of trust. We outline five broad use cases where explanations are useful and, for each, we describe pseudo-experiments that rely on objective empirical measurements and falsifiable hypotheses. We believe that this experimental rigor is necessary to contribute to scientific knowledge in the field of XAI.

Mobile edge computing (MEC), an emerging technology, has the characteristics of low latency, mobile energy savings, and context-awareness. As a type of access network, wireless mesh network (WMN) has gained wide attention due to its flexible network architecture, low deployment cost, and self-organization. The combination of MEC and WMN can solve the shortcomings of traditional wireless communication such as storage capacity, privacy, and security. In this paper, we propose a location-aware (LA) algorithm to cognize the location and a location-aware offloading policy (LAOP) algorithm considering the energy consumption and time delay. Simulation results show that the proposed LAOP algorithm can obtain a higher completion rate and lower average processing delay compared with the other two methods.

The existing network intrusion detection methods have less label samples in the training process, and the detection accuracy is not high. In order to solve this problem, this paper designs a network intrusion detection method based on the GAN model by using the adversarial idea contained in the GAN. The model enhances the original training set by continuously generating samples, which expanding the label sample set. In order to realize the multi-classification of samples, this paper transforms the previous binary classification model of the generated adversarial network into a supervised learning multi-classification model. The loss function of training is redefined, so that the corresponding training method and parameter setting are obtained. Under the same experimental conditions, several performance indicators are used to compare the detection ability of the proposed method, the original classification model and other models. The experimental results show that the method proposed in this paper is more stable, robust, accurate detection rate, has good generalization ability, and can effectively realize network intrusion detection.

This paper showcases multiclass classification baselines using different machine learning algorithms and neural networks for distinguishing legitimate network traffic from direct and obfuscated network intrusions. This research derives its baselines from Advanced Security Network Metrics & Tunneling Obfuscations dataset. The dataset captured legitimate and obfuscated malicious TCP communications on selected vulnerable network services. The multiclass classification NIDS is able to distinguish obfuscated and direct network intrusion with up to 95% accuracy.

To accommodate the rapidly changing Internet requirements, Information-Centric Networking (ICN) was recently introduced as a promising architecture for the future Internet. One of the ICN primary features is `in-network caching'; due to its ability to minimize network traffic and respond faster to users' requests. Therefore, various caching algorithms have been presented that aim to enhance the network performance using different measures, such as cache hit ratio and cache hit distance. Choosing a caching strategy is critical, and an adequate replacement strategy is also required to decide which content should be dropped. Thus, in this paper, we propose a content replacement scheme for ICN, called Randomized LFU that is implemented with respect to content popularity taking the time complexity into account. We use Abilene and Tree network topologies in our simulation models. The proposed replacement achieves encouraging results in terms of the cache hit ratio, inner hit, and hit distance and it outperforms FIFO, LRU, and Random replacement strategies.

Developing mission-centric impact assessment techniques to address cyber resiliency in the cyber-physical systems (CPSs) requires integrating system inter-dependencies to the risk and resilience analysis process. Generally, network administrators utilize attack graphs to estimate possible consequences in a networked environment. Attack graphs lack to incorporate the operations-specific dependencies. Localizing the dependencies among operational missions, tasks, and the hosting devices in a large-scale CPS is also challenging. In this work, we offer a graphical modeling technique to integrate the mission-centric impact assessment of cyberattacks by relating the effect to the operational resiliency by utilizing a combination of the logical attack graph and mission impact propagation graph. We propose formal techniques to compute cyberattacks’ impact on the operational mission and offer an optimization process to minimize the same, having budgetary restrictions. We also relate the effect to the system functional operability. We illustrate our modeling techniques using a SCADA (supervisory control and data acquisition) case study for the cyber-physical power systems. We believe our proposed method would help evaluate and minimize the impact of cyber attacks on CPS’s operational missions and, thus, enhance cyber resiliency.

As any other infrastructure relying on cyber-physical systems (CPS), naval CPS are highly interconnected and collect considerable data streams, on which depend multiple command and navigation decisions. Being a data-driven decision system requiring optimized supervisory control on a permanent basis, it is critical to examine the CPS vulnerability to anomalies and their propagation. This paper presents an approach to detect CPS anomalies and estimate their propagation applying a quality assessed graph, which represents the CPS physical and digital subsystems, combined with system variables dependencies and a set of data and information quality measures vectors. Following the identification of variables dependencies and high-risk nodes in the CPS, data and information quality measures reveal how system variables are modified when an anomaly is detected, also indicating its propagation path. Taking as reference the normal state of a naval propulsion management system, four anomalies in the form of cyber-attacks - port scan, programmable logical controller stop, and man in the middle to change the motor speed and operation of a tank valve - were produced. Three anomalies were properly detected and their propagation path identified. These results suggest the feasibility of anomaly detection and estimation of propagation estimation in CPS, applying data and information quality analysis to a system graph.

Aiming at the problem that the traditional intrusion detection method can not effectively deal with the massive and high-dimensional network traffic data of industrial control system (ICS), an ICS intrusion detection strategy based on bidirectional generative adversarial network (BiGAN) is proposed in this paper. In order to improve the applicability of BiGAN model in ICS intrusion detection, the optimal model was obtained through the single variable principle and cross-validation. On this basis, the supervised control and data acquisition (SCADA) standard data set is used for comparative experiments to verify the performance of the optimized model on ICS intrusion detection. The results show that the ICS intrusion detection method based on optimized BiGAN has higher accuracy and shorter detection time than other methods.

For real-time applications of arbitrary style transformation, there is a trade-off between the quality of results and the running time of existing algorithms. Hence, it is required to maintain the equilibrium of the quality of generated artwork with the speed of execution. It's complicated for the present arbitrary style-transformation procedures to preserve the structure of content-image while blending with the design and pattern of style-image. This paper presents the implementation of a network using SANET models for generating impressive artworks. It is flexible in the fusion of new style characteristics while sustaining the semantic-structure of the content-image. The identity-loss function helps to minimize the overall loss and conserves the spatial-arrangement of content. The results demonstrate that this method is practically efficient, and therefore it can be employed for real-time fusion and transformation using arbitrary styles.

Deep Neural Networks (DNNs) are susceptible to model stealing attacks, which allows a data-limited adversary with no knowledge of the training dataset to clone the functionality of a target model, just by using black-box query access. Such attacks are typically carried out by querying the target model using inputs that are synthetically generated or sampled from a surrogate dataset to construct a labeled dataset. The adversary can use this labeled dataset to train a clone model, which achieves a classification accuracy comparable to that of the target model. We propose "Adaptive Misinformation" to defend against such model stealing attacks. We identify that all existing model stealing attacks invariably query the target model with Out-Of-Distribution (OOD) inputs. By selectively sending incorrect predictions for OOD queries, our defense substantially degrades the accuracy of the attacker's clone model (by up to 40%), while minimally impacting the accuracy (\textbackslashtextless; 0.5%) for benign users. Compared to existing defenses, our defense has a significantly better security vs accuracy trade-off and incurs minimal computational overhead.

Securing cyber-physical systems (CPS) and Internet of Things (IoT) systems requires the identification of how interdependence among existing atomic vulnerabilities may be exploited by an adversary to stitch together an attack that can compromise the system. Therefore, accurate attack graphs play a significant role in systems security. A manual construction of the attack graphs is tedious and error-prone, this paper proposes a model-checking-based automated attack graph generator and visualizer (A2G2V). The proposed A2G2V algorithm uses existing model-checking tools, an architecture description tool, and our own code to generate an attack graph that enumerates the set of all possible sequences in which atomic-level vulnerabilities can be exploited to compromise system security. The architecture description tool captures a formal representation of the networked system, its atomic vulnerabilities, their pre-and post-conditions, and security property of interest. A model-checker is employed to automatically identify an attack sequence in the form of a counterexample. Our own code integrated with the model-checker parses the counterexamples, encodes those for specification relaxation, and iterates until all attack sequences are revealed. Finally, a visualization tool has also been incorporated with A2G2V to generate a graphical representation of the generated attack graph. The results are illustrated through application to computer as well as control (SCADA) networks.

Autonomy in cybersystems depends on their ability to be self-aware by understanding the intent of services and applications that are running on those systems. In case of mission-critical cybersystems that are deployed in dynamic and unpredictable environments, the newly integrated unknown applications or services can either be benign and essential for the mission or they can be cyberattacks. In some cases, these cyberattacks are evasive Advanced Persistent Threats (APTs) where the attackers remain undetected for reconnaissance in order to ascertain system features for an attack e.g. Trojan Laziok. In other cases, the attackers can use the system only for computing e.g. cryptomining malware. APTs such as cryptomining malware neither disrupt normal system functionalities nor trigger any warning signs because they simply perform bitwise and cryptographic operations as any other benign compression or encoding application. Thus, it is difficult for defense mechanisms such as antivirus applications to detect these attacks. In this paper, we propose an Operating Context profiling system based on deep neural networks-Long Short-Term Memory (LSTM) networks-using Windows Performance Counters data for detecting these evasive cryptomining applications. In addition, we propose Deep Cryptomining Profiler (DeCrypto Pro), a detection system with a novel model selection framework containing a utility function that can select a classification model for behavior profiling from both the light-weight machine learning models (Random Forest and k-Nearest Neighbors) and a deep learning model (LSTM), depending on available computing resources. Given data from performance counters, we show that individual models perform with high accuracy and can be trained with limited training data. We also show that the DeCrypto Profiler framework reduces the use of computational resources and accurately detects cryptomining applications by selecting an appropriate model, given the constraints such as data sample size and system configuration.

Advanced Persistent Threats (APTs) are stealthy, sophisticated, long-term, multi-stage attacks that threaten the security of sensitive information. Dynamic Information Flow Tracking (DIFT) has been proposed as a promising mechanism to detect and prevent various cyber attacks in computer systems. DIFT tracks suspicious information flows in the system and generates security analysis when anomalous behavior is detected. The number of information flows in a system is typically large and the amount of resources (such as memory, processing power and storage) required for analyzing different flows at different system locations varies. Hence, efficient use of resources is essential to maintain an acceptable level of system performance when using DIFT. On the other hand, the quickest detection of APTs is crucial as APTs are persistent and the damage caused to the system is more when the attacker spends more time in the system. We address the problem of detecting APTs and model the trade-off between resource efficiency and quickest detection of APTs. We propose a game model that captures the interaction of APT and a DIFT-based defender as a two-player, multi-stage, zero-sum, Stackelberg semi-Markov game. Our game considers the performance parameters such as false-negatives generated by DIFT and the time required for executing various operations in the system. We propose a two-time scale Q-learning algorithm that converges to a Stackelberg equilibrium under infinite horizon, limiting average payoff criteria. We validate our model and algorithm on a real-word attack dataset obtained using Refinable Attack INvestigation (RAIN) framework.

It is necessary to improve the safety of the underwater acoustic sensor networks (UASNs) since it is mostly used in the military industry. Specific emitter identification is the process of identifying different transmitters based on the radio frequency fingerprint extracted from the received signal. The sonar transmitter is a typical low-frequency radiation source and is an important part of the UASNs. Class D power amplifier, a typical nonlinear amplifier, is usually used in sonar transmitters. The inherent nonlinearity of power amplifiers provides fingerprint features that can be distinguished without transmitters for specific emitter recognition. First, the nonlinearity of the sonar transmitter is studied in-depth, and the nonlinearity of the power amplifier is modeled and its nonlinearity characteristics are analyzed. After obtaining the nonlinear model of an amplifier, a similar amplifier in practical application is obtained by changing its model parameters as the research object. The output signals are collected by giving the same input of different models, and, then, the output signals are extracted and classified. In this paper, the memory polynomial model is used to model the amplifier. The power spectrum features of the output signals are extracted as fingerprint features. Then, the dimensionality of the high-dimensional features is reduced. Finally, the classifier is used to recognize the amplifier. The experimental results show that the individual sonar transmitter can be well identified by using the nonlinear characteristics of the signal. By this way, this method can enhance the communication safety of the UASNs.

In this paper, we apply machine learning to distributed private data owned by multiple data owners, entities with access to non-overlapping training datasets. We use noisy, differentially-private gradients to minimize the fitness cost of the machine learning model using stochastic gradient descent. We quantify the quality of the trained model, using the fitness cost, as a function of privacy budget and size of the distributed datasets to capture the trade-off between privacy and utility in machine learning. This way, we can predict the outcome of collaboration among privacy-aware data owners prior to executing potentially computationally-expensive machine learning algorithms. Particularly, we show that the difference between the fitness of the trained machine learning model using differentially-private gradient queries and the fitness of the trained machine model in the absence of any privacy concerns is inversely proportional to the size of the training datasets squared and the privacy budget squared. We successfully validate the performance prediction with the actual performance of the proposed privacy-aware learning algorithms, applied to: financial datasets for determining interest rates of loans using regression; and detecting credit card frauds using support vector machines.

The limited information on the cyberattacks available in the unclassified regime, hardens standardizing the analysis. We address the problem of modeling and analyzing cyberattacks using a multimodal graph approach. We formulate the stages, actors, and outcomes of cyberattacks as a multimodal graph. Multimodal graph nodes include cyberattack victims, adversaries, autonomous systems, and the observed cyber events. In multimodal graphs, single-modality graphs are interconnected according to their interaction. We apply community and centrality analysis on the graph to obtain in-depth insights into the attack. In community analysis, we cluster those nodes that exhibit “strong” inter-modal ties. We further use centrality to rank the nodes according to their importance. Classifying nodes according to centrality provides the progression of the attack from the attacker to the targeted nodes. We apply our methods to two popular case studies, namely GhostNet and Putter Panda and demonstrate a clear distinction in the attack stages.

With the increasing performance of neural networks on many security-critical tasks, the security concerns of machine learning have become increasingly prominent. Recent studies have shown that neural networks are vulnerable to adversarial examples: carefully crafted inputs with negligible perturbations on legitimate samples could mislead a neural network to produce adversary-selected outputs while humans can still correctly classify them. Therefore, we need an additional measurement on the trustworthiness of the results of a machine learning model, especially in adversarial settings. In this paper, we analyse the root cause of adversarial examples, and propose a new property, namely fidelity, of machine learning models to describe the gap between what a model learns and the ground truth learned by humans. One of its benefits is detecting adversarial attacks. We formally define fidelity, and propose a novel approach to quantify it. We evaluate the quantification of fidelity in adversarial settings on two neural networks. The study shows that involving the fidelity enables a neural network system to detect adversarial examples with true positive rate 97.7%, and false positive rate 1.67% on a studied neural network.

In service-oriented computing environment (e.g. cloud computing), Cloud Customers (CCs) and Cloud Service Providers (CSPs) require to calculate the trust ranks of impending partner prior to appealing in communications. Determining trustworthiness dynamically is a demanding dilemma in an open and dynamic environment (such as cloud computing) because of many CSPs providing same types of services. Presently, there are very less number of dynamic trust evaluation scheme that permits CCs to evaluate CSPs trustworthiness from multi-dimensional perspectives. Similarly, there is no scheme that permits CSPs to evaluate trustworthiness of CCs. This paper proposes a Multidimensional Dynamic Trust Evaluation Scheme (MDTES) that facilitates CCs to evaluate the trustworthiness of CSPs from various viewpoints. Similar approach can be employed by CSPs to evaluate the trustworthiness of CCs. The proposed MDTES helps CCs to choose trustworthy CSP and to have desired QoS requirements and CSPs to choose desired and legal CCs. The simulation results illustrate the MDTES is dynamic and steady in distinguishing trustworthy and untrustworthy CSPs and CCs.

In the Internet of Things architecture, devices are frequently connected to the Internet either directly or indirectly. However, many IoT devices lack built-in security features such as device level encryption, user authentication and basic firewall protection. This paper discusses security risks in the layers of general Internet of Things architecture and shows examples of potential risks at each level of the architecture. The paper also compares IoT security solutions provided by three major vendors and shows that the solutions are mutually complementary. Nevertheless, none of the examined IoT solutions provides security at the device level of the IoT architecture model. In order to address risks at the device level of the architecture, an implementation of Trusted Platform Module and Unique Device Identifier on IoT devices and gateways for encryption, authentication and device management is advocated in the paper.

Verifying complex Cyber-Physical Systems (CPS) is increasingly important given the push to deploy safety-critical autonomous features. Unfortunately, traditional verification methods do not scale to the complexity of these systems and do not provide systematic methods to protect verified properties when not all the components can be verified. To address these challenges, this paper proposes a real-time mixed-trust computing framework that combines verification and protection. The framework introduces a new task model, where an application task can have both an untrusted and a trusted part. The untrusted part allows complex computations supported by a full OS with a realtime scheduler running in a VM hosted by a trusted hypervisor. The trusted part is executed by another scheduler within the hypervisor and is thus protected from the untrusted part. If the untrusted part fails to finish by a specific time, the trusted part is activated to preserve safety (e.g., prevent a crash) including its timing guarantees. This framework is the first allowing the use of untrusted components for CPS critical functions while preserving logical and timing guarantees, even in the presence of malicious attackers. We present the framework design and implementation along with the schedulability analysis and the coordination protocol between the trusted and untrusted parts. We also present our Raspberry Pi 3 implementation along with experiments showing the behavior of the system under failures of untrusted components, and a drone application to demonstrate its practicality.

This paper aims at identifying the neural correlates of human trust in autonomous systems using electroencephalography (EEG) signals. Quantifying the relationship between trust and brain activities allows for real-time assessment of human trust in automation. This line of effort contributes to the design of trusted autonomous systems, and more generally, modeling the interaction in human-autonomy interaction. To study the correlates of trust, we use an investment game in which artificial agents with different levels of trustworthiness are employed. We collected EEG signals from 10 human subjects while they are playing the game; then computed three types of features from these signals considering the signal time-dependency, complexity and power spectrum using an autoregressive model (AR), sample entropy and Fourier analysis, respectively. Results of a mixed model analysis showed significant correlation between human trust and EEG features from certain electrodes. The frontal and the occipital area are identified as the predominant brain areas correlated with trust.

Computational trust is the digital counterpart of the human notion of trust as applied in social systems. Its main purpose is to improve the reliability of interactions in online communities and of knowledge transfer in information management systems. Trust models are formal frameworks in which the notion of computational trust is described rigorously and where its dynamics are explained precisely. In this paper we will consider and extend a computational trust model, i.e., JØsang's Subjective Logic: we will show how this model is well-suited to describe the dynamics of computational trust, but lacks effective tools to compute initial trust values to feed in the model. To overcome some of the issues with subjective logic, we will introduce a logical language which can be employed to describe and reason about trust. The core ideas behind the logical language will turn out to be useful in computing initial trust values to feed into subjective logic. The aim of the paper is, therefore, that of providing an improvement on subjective logic.

Evaluating user behavior in cloud computing infrastructure is important for both Cloud Users and Cloud Service Providers. The service providers must ensure the safety of users who access the cloud. User behavior can be modeled and employed to help assess trust and play a role in ensuring authenticity and safety of the user. In this paper, we propose a User Behavior Trust Model based on Fuzzy Logic (UBTMFL). In this model, we develop user history patterns and compare them current user behavior. The outcome of the comparison is sent to a trust computation center to calculate a user trust value. This model considers three types of trust: direct, history and comprehensive. Simulation results are included.

Nowadays, with the diversification and decentralization of energy systems, the energy Internet makes it possible to interconnect distributed energy sources and consumers. In the energy trading market, the traditional centralized model relies entirely on trusted third parties. However, as the number of entities involved in the transactions grows and the forms of transactions diversify, the centralized model gradually exposes problems such as insufficient scalability, High energy consumption, and low processing efficiency. To address these challenges, we propose a secure and efficient energy renewable trading scheme based on blockchain. In our scheme, the electricity market trading model is divided into two levels, which can not only protect the privacy, but also achieve a green computing. In addition, in order to adapt to the relatively weak computing power of the underlying equipment in smart grid, we design a credibility-based equity proof mechanism to greatly improve the system availability. Compared with other similar distributed energy trading schemes, we prove the advantages of our scheme in terms of high operational efficiency and low computational overhead through experimental evaluations. Additionally, we conduct a detailed security analysis to demonstrate that our solution meets the security requirements.