Biblio

Machine Learning (ML) models are now commonly used as components in systems. As any other component, ML components can produce erroneous outputs that may penalize system utility. In this context, self-adaptive systems emerge as a natural approach to cope with ML mispredictions, through the execution of adaptation tactics such as model retraining. To synthesize an adaptation strategy, the self-adaptation manager needs to reason about the cost-benefit tradeoffs of the applicable tactics, which is a non-trivial task for tactics such as model retraining, whose benefits are both context- and data-dependent.To address this challenge, this paper proposes a probabilistic modeling framework that supports automated reasoning about the cost/benefit tradeoffs associated with improving ML components of ML-based systems. The key idea of the proposed approach is to decouple the problems of (i) estimating the expected performance improvement after retrain and (ii) estimating the impact of ML improved predictions on overall system utility.We demonstrate the application of the proposed framework by using it to self-adapt a state-of-the-art ML-based fraud-detection system, which we evaluate using a publicly-available, real fraud detection dataset. We show that by predicting system utility stemming from retraining a ML component, the probabilistic model checker can generate adaptation strategies that are significantly closer to the optimal, as compared against baselines such as periodic retraining, or reactive retraining.

Website phishing continues to persist as one of the most important security threats of the modern Internet era. A major concern has been that machine learning based approaches, which have been the cornerstones of deployed phishing detection solutions, have not been able to adapt to the evolving nature of the phishing attacks. To create updated machine learning models, the collection of a sufficient corpus of real-time phishing data has always been a challenging problem as most phishing websites are short-lived. In this work, for the first time, we address these important concerns and describe an adaptive phishing detection solution that is able to adapt to changes in phishing attacks. Our solution has two major contributions. First, our solution allows for multiple organizations to collaborate in a privacy preserving manner and generate a robust machine learning model for phishing detection. Second, our solution is designed to be flexible in order to adapt to the novel phishing features introduced by attackers. Our solution not only allows for incorporating novel features into the existing machine learning model, but also can help, to a certain extent, the “unlearning” of existing features that have become obsolete in current phishing attacks. We evaluated our approach on a large real-world data collected over a period of six months. Our results achieve a high true positive rate of 97 %, which is on par with existing state-of-the art centralized solutions. Importantly, our results demonstrate that, a machine learning model can incorporate new features while selectively “unlearning” the older obsolete features.

Smart city technologies promise reduced congestion by optimizing transportation movements. Increased connectivity, however, may increase the attack surface of a municipality's critical functions. Increased supply chain attacks (up nearly 80 % in 2019) and municipal ransomware attacks (up 60 % in 2019) motivate the need for holistic approaches to risk assessment. Therefore, we present a methodology to quantify the degree to which supply-chain movements may be observed or disrupted via compromised smart-city devices. Our data-processing pipeline uses publicly available datasets to model intermodal commodity flows within and surrounding a municipality. Using a hierarchy tree to adaptively sample spatial networks within geographic regions of interest, we bridge the gap between grid- and network-based risk assessment frameworks. Results based on fieldwork for the Jack Voltaic exercises sponsored by the Army Cyber Institute demonstrate our approach on intermodal movements through Charleston, SC and San Diego, CA.

Negotiation is a complex social interaction that encapsulates emotional encounters in human decision-making. Virtual agents that can negotiate with humans are useful in pedagogy and conversational AI. To advance the development of such agents, we explore the prediction of two important subjective goals in a negotiation – outcome satisfaction and partner perception. Specifically, we analyze the extent to which emotion attributes extracted from the negotiation help in the prediction, above and beyond the individual difference variables. We focus on a recent dataset in chat-based negotiations, grounded in a realistic camping scenario. We study three degrees of emotion dimensions – emoticons, lexical, and contextual by leveraging affective lexicons and a state-of-the-art deep learning architecture. Our insights will be helpful in designing adaptive negotiation agents that interact through realistic communication interfaces.

Compared with other virtualization technologies, container technology is widely used in edge computing because of its low cost, high reliability, high flexibility and fast portability. However, the use of container technology can alleviate the pressure of massive data, but also bring complex and diverse security problems. Reliable information security risk assessment method is the key to ensure the smooth application of container technology. According to the risk assessment theory, a security risk assessment method for edge computing containers based on dynamic game theory is proposed. Aiming at the complex container security attack and defense process, the container system's security model is constructed based on dynamic game theory. By combining the attack and defense matrix, the Nash equilibrium solution of the model is calculated, and the dynamic process of the mutual game between security defense and malicious attackers is analyzed. By solving the feedback Nash equilibrium solution of the model, the optimal strategies of the attackers are calculated. Finally, the simulation tool is used to solve the feedback Nash equilibrium solution of the two players in the proposed model, and the experimental environment verifies the usability of the risk assessment method.

As a supplement to the cloud computing model, the edge computing model can use edge servers and edge devices to coordinate information processing on the edge of the network to help Internet of Thing (IoT) data storage, transmission, and computing tasks. In view of the complex and changeable situation of edge computing IoT scenarios, this paper proposes a multi-dimensional trust evaluation factor selection scheme. Improve the traditional trusted modeling method based on direct/indirect trust, introduce multi-dimensional trusted decision attributes and rely on the collaboration of edge servers and edge device nodes to infer and quantify the trusted relationship between nodes, and combine the information entropy theory to smoothly weight the calculation results of multi-dimensional decision attributes. Improving the current situation where the traditional trusted assessment scheme's dynamic adaptability to the environment and the lack of reliability of trusted assessment are relatively lacking. Simulation experiments show that the edge computing IoT multi-dimensional trust evaluation model proposed in this paper has better performance than the trusted model in related literature.

In the last decade the use of artificial neural networks (ANNs) in many fields like image processing or speech recognition has become a common practice because of their effectiveness to solve complex tasks. However, in such a rush, very little attention has been paid to security aspects. In this work we explore the possibility to embed a watermark into the ANN parameters. We exploit model redundancy and adaptation capacity to lock a subset of its parameters to carry the watermark sequence. The watermark can be extracted in a simple way to claim copyright on models but can be very easily attacked with model fine-tuning. To tackle this culprit we devise a novel watermark aware training strategy. We aim at delving into the loss landscape to find an optimal configuration of the parameters such that we are robust to fine-tuning attacks towards the watermarked parameters. Our experimental results on classical ANN models trained on well-known MNIST and CIFAR-10 datasets show that the proposed approach makes the embedded watermark robust to fine-tuning and compression attacks.

Crypto-ransomware is a malware that uses the system’s cryptography functions to encrypt user data. The irreversible effect of crypto-ransomware makes it challenging to survive the attack compared to other malware categories. When a crypto-ransomware attack encrypts user files, it becomes difficult to access these files without having the decryption key. Due to the availability of ransomware development tool kits like Ransomware as a Service (RaaS), many ransomware variants are being developed. This contributes to the rise of ransomware attacks witnessed nowadays. However, the conventional approaches employed by malware detection solutions are not suitable to detect ransomware. This is because ransomware needs to be detected as early as before the encryption takes place. These attacks can effectively be handled only if detected during the pre-encryption phase. Early detection of ransomware attacks is challenging due to the limited amount of data available before encryption. An adaptive pre-encryption model is proposed in this paper which is expected to deal with the population concept drift of crypto-ransomware given the limited amount of data collected during the pre-encryption phase of the attack lifecycle. With such adaptability, the model can maintain up-to-date knowledge about the attack behavior and identify the polymorphic ransomware that continuously changes its behavior.

With the development of artificial intelligence and cryptography, intelligent cryptography will be the trend of encrypted communications in the future. Abadi designed an encrypted communication model based on a generative adversarial network, which can communicate securely when the adversary knows the ciphertext. The communication party and the adversary fight against each other to continuously improve their own capabilities to achieve a state of secure communication. However, this model can only have a better communication effect under the 16 bits communication length, and cannot adapt to the length of modern encrypted communication. Combine the neural network structure in DCGAN to optimize the neural network of the original model, and at the same time increase the batch normalization process, and optimize the loss function in the original model. Experiments show that under the condition of the maximum 2048-bit communication length, the decryption success rate of communication reaches about 0.97, while ensuring that the adversary’s guess error rate is about 0.95, and the training speed is greatly increased to keep it below 5000 steps, ensuring safety and efficiency Communication.

In human-machine systems (HMS), trust placed by humans on machines is a complex concept and attracts increasingly research efforts. Herein, we reviewed recent studies on building and measuring trust in HMS. The review was based on one comprehensive model of trust – IMPACTS, which has 7 features of intention, measurability, performance, adaptivity, communication, transparency, and security. The review found that, in the past 5 years, HMS fulfill the features of intention, measurability, communication, and transparency. Most of the HMS consider the feature of performance. However, all of the HMS address rarely the feature of adaptivity and neglect the feature of security due to using stand-alone simulations. These findings indicate that future work considering the features of adaptivity and/or security is imperative to foster human trust in HMS.

Adaptive honeypots are being widely proposed as a more powerful alternative to the traditional honeypot model. Just as with typical honeypots, however, one of the most important concerns of an adaptive honeypot is environment deception in order to make sure an adversary cannot fingerprint the honeypot. The threat of fingerprinting hints at a greater underlying concern, however; this being that honeypots are only effective because an adversary does not know that the environment on which they are operating is a honeypot. What has not been widely discussed in the context of adaptive honeypots is that they actually have an inherently increased level of susceptibility to this threat. Honeypots not only bear increased risks when an adversary knows they are a honeypot rather than a native system, but they are only effective as adaptable entities if one does not know that the honeypot environment they are operating on is adaptive as wekk. Thus, if adaptive honeypots become commonplace - or, instead, if attackers even have an inkling that an adaptive honeypot may exist on any given network, a new attack which could develop is a “blind confusion attack”; a form of connection which simply makes an assumption all environments are adaptive honeypots, and instead of attempting to perform a malicious strike on a given entity, opts to perform non-malicious behavior in specified and/or random patterns to confuse an adaptive network's learning.

The full integration of Remotely Piloted Aircraft Systems (RPAS) in non-segregated airspace is one of the major objectives for the worldwide aviation organizations and authorities. However, there are several technological and regulatory issues due to the increase of the air traffic in the next years and to the need of keeping high safety levels. In this framework, a real-time validation environment capable to simulate complex scenarios related to future air traffic management (ATM) conditions is of paramount importance. These facilities allow detailed testing and tuning of new technologies and procedures before executing flight tests. With such motivations, the Italian Aerospace Research Centre has developed the Integrated Simulation Facility (ISF) able to accurately reproduce ATM complex scenarios in real-time with hardware and human in-the-loop simulations, aiming to validate new ATM procedures and innovative system prototypes for RPAS and General Aviation aircraft. In the present work, the ISF facility has been used for reproducing relevant ATM scenarios to validate the functionalities of a Detect and Avoid system (DAA). The results of the ISF test campaign demonstrate the effectiveness of the developed algorithm in the autonomous resolution of mid-air collisions in presence of both air traffic and fixed obstacles (i.e. bad weather areas, no-fly-zone and terrain) and during critical flight phases, thus exceeding the current DAA state-of-the-art.

This work presents a system which allows image data and extracted features from a real-world location to be captured and modelled in a Virtual Reality (VR) environment combined with Augmented Situated Visualizations (ASV) overlaid and registered in a virtual environment. Combining these technologies with techniques from Data Science and Artificial Intelligence (AI)(such as image analysis and 3D reconstruction) allows the creation of a setting where remote locations can be modelled and interacted with from anywhere in the world. This Enhanced Photorealistic Immersive (EPI) system is highly adaptable to a wide range of use cases and users as it can be utilized to model and interact with any environment which can be captured as image data (such as training for operation in hazardous environments, accessibility solutions for exploration of historical/tourism locations and collaborative learning environments). A use case example focused on a structural examination of railway tunnels along with a pilot study is presented, which can demonstrate the usefulness of the EPI system.

This research attempts to provide some insights on the application of text mining and Natural Language Processing (NLP). The application domain is consumer complaints about financial institutions in the USA. As an advanced analytics discipline embedded within the Big Data paradigm, the practice of text analytics contains elements of emergent knowledge processes. Since our experiment should be able to scale up we make use of a pipeline based on Spark-NLP. The usage scenario is adapting the model to a specific industrial context and using the dataset offered by the "Consumer Financial Protection Bureau" to illustrate the application.

With the popularity of smart devices in the power grid and the advancement of data collection technology, the amount of electricity usage data has exploded in recent years, which is beneficial for optimizing service quality and grid operation. However, current data analysis is mainly based on cloud platforms, which poses challenges to transmission bandwidth, computing resources, and transmission delays. To solve the problem, this paper proposes a graph convolution neural networks (GCNs) based edge-cloud collaborative anomaly detection model. Specifically, the time series is converted into graph data based on visibility graph model, and graph convolutional network model is adopted to classify the labeled graph data for anomaly detection. Then a model segmentation method is proposed to adaptively divide the anomaly detection model between the edge equipment and the back-end server. Experimental results show that the proposed scheme provides an effective solution to edge anomaly detection and can make full use of the computing resources of terminal equipment.

This paper presents ongoing work on the formalism of Cyber-Physical Systems (CPS) simulations. These systems are distributed real-time systems, and their simulations might be distributed or not. In this paper, we propose a model to describe the modular components forming a simulation of a CPS. The main goal is to introduce a model of generic simulation distributed architecture, on which we are able to execute a logical architecture of simulation. This architecture of simulation allows the expression of structural and behavioural constraints on the simulation, abstracting its execution. We will propose two implementations of the execution architecture based on generic architectures of distributed simulation: $\cdot$ The High Level Architecture (HLA), an IEEE standard for distributed simulation, and one of its open-source implementation of RunTime Infrastructure (RTI): CERTI. $\cdot$ The Distributed Simulation Scheduler (DSS), an Airbus framework scheduling predefined models. Finally, we present the initial results obtained applying our formalism to the open-source case study from the ROSACE case study.

Distributed acoustic sensing (DAS) systems based on fiber brag grating (FBG) have been widely used for distributed temperature and strain sensing over the past years, and function well in perimeter security monitoring and structural health monitoring. However, with relevant algorithms functioning with low accuracy, the DAS system presently has trouble in signal recognition, which puts forward a higher requirement on a high-precision identification method. In this paper, we propose an improved recognition method based on relative fundamental signal processing methods and convolutional neural network (CNN) to construct a mathematical model of disturbance FBG signal recognition. Firstly, we apply short-time energy (STE) to extract original disturbance signals. Secondly, we adopt short-time Fourier transform (STFT) to divide a longer time signal into short segments. Finally, we employ a CNN model, which has already been trained to recognize disturbance signals. Experimental results conducted in the real environments show that our proposed algorithm can obtain accuracy over 96.5%.

In this paper, we present a theoretical approach concerning the econometric modelling for the estimation of cyber-security risk, with the use of time-series analysis methods and alternatively with Machine Learning (ML) based, deep learning methodology. Also we present work performed in the framework of SAINT H2020 Project [1], concerning innovative data mining techniques, based on automated web scrapping, for the retrieving of the relevant time-series data. We conclude with a review of emerging challenges in cyber-risk assessment brought by the rapid development of adversarial AI.

Easy Data Augmentation is originally developed for text classification tasks. It consists of four basic methods: Synonym Replacement, Random Insertion, Random Deletion, and Random Swap. They yield accuracy improvements on several deep neural network models. In this study we apply these methods to a new domain. We augment Named Entity Recognition datasets from medical domain. Although the augmentation task is much more difficult due to the nature of named entities which consist of word or word groups in the sentences, we show that we can improve the named entity recognition performance.

The need to achieve a suitable level of security in Cyber-Physical Systems (CPS) presents a major challenge for engineers. The unpredictable communication of highly constrained, but safety-relevant systems in a heterogeneous environment, significantly impacts the number and severity of vulnerabilities. Consequently, if security-related weaknesses can successfully be exploited by attackers, the functionality of critical infrastructure could be denied or malfunction. This might consequently threaten life or leak sensitive information. A toolkit to quantitatively express security is essential for security engineers in order to define security-enhancing measurements. For this purpose, security scoring frameworks, like the established Common Vulnerability Scoring System can be used. However, existing security scoring frameworks may not be able to handle the proposed challenges and characteristics of CPS. Therefore, in this work, we aim to elaborate a security scoring system that is tailored to the needs of CPS. In detail, we analyze security on a System-of-Systems level, while considering multiple attacks, as well as potential side effects to other security-related objects. The positive effects of integrated mitigation concepts should also be abbreviated by our proposed security score. Additionally, we generate the security score for interacting AUTOSAR platforms in a highly-connected Vehicle-to-everything (V2x) environment. We refer to this highly relevant use case scenario to underline the benefits of our proposed scoring framework and to prove its effectiveness in CPS.

One of the measures to prevent multi-pass social engineering attacks is to identify the chains of user, which are most susceptible to such attacks. The aim of the study is to combine a mathematical model for estimating the probability of success of the propagation of a multi-pass social engineering attack between users with a model for calculating information influence. Namely, it is proposed to include in estimating the intensity of interactions between users (which used in the model of the propagation of a multi-pass social engineering attack) estimating of power of influence actions of agents. The scientific significance of the work consists in the development of a mathematical structure for modeling the actions of an attacker-social engineer and creating a foundation for the subsequent analysis of the social graph of the organization's employees. The practical significance lies in the formation of opportunities for decision-makers. Therefore, they will be able to take more precise measures for increase the level of security as individual employees as the organization generally.

In order to increase the availability and reliability of photovoltaic (PV) systems, fault diagnosis and condition monitoring of inverters are of crucial means to meet the goals. Numerous methods are implemented for fault diagnosis of PV inverters, providing robust features and handling massive amount of data. However, existing methods rely on simplistic frameworks that are incapable of inspecting a wide range of intrinsic and explicit features, as well as being time-consuming. In this paper, a novel method based on a multilayer deep belief network (DBN) is suggested for fault diagnosis, which allows the framework to discover the probabilistic reconstruction across its inputs. This approach equips a robust hierarchical generative model for exploiting features associated with faults, interprets functions that are highly variable, and needs lesser prior information. Moreover, the method instantaneously categorizes the fault conditions, which eventually strengthens the adaptability of applying it on a variety of diagnostic problems in an inverter domain. The proposed method is evaluated using multiple input signals at different sampling frequencies. To evaluate the efficacy of DBN, a test model based on a three-phase 2-level grid-tied PV inverter was used. The results show that the method is capable of achieving precise diagnosis operations.

The model-free adaptive security tracking control (MFASTC) problem of nonlinear networked control systems is explored in this paper with DoS attacks and delays consideration. In order to alleviate the impact of DoS attack and RTT delays on NCSs performance, an attack compensation mechanism and a networked predictive-based delay compensation mechanism are designed, respectively. The data-based designed method need not the dynamic and structure of the system, The MFASTC algorithm is proposed to ensure the output tracking error being bounded in the mean-square sense. Finally, an example is given to illustrate the effectiveness of the new algorithm by a comparison.

In this paper, we propose models and schemes for coded and uncoded packet transmission over time invariant (TIC) and time variant (TVC) channels. We provide an approximation of the delay induced assuming fmite number of time slots to transmit a given number of packets. We propose an adaptive physical layer (PHY)-aware coded scheme that designs smart acknowledgments (ACK) via an optimal selection of coded packets to transmit at a given SNR. We apply our proposed schemes to channels with complex fading behavior and high round trip (RTT) delays. We compare the accuracy of TVC coded scheme to the TIC coded scheme, and we show the throughput-delay efficacy of adaptive coded schemes driven by PHY-awareness in the mitigation of high RTT environments, with up to 3 fold gains.

This work seeks to advance the state of the art in HPC I/O performance analysis and interpretation. In particular, we demonstrate effective techniques to: (1) model output performance in the presence of I/O interference from production loads; (2) build features from write patterns and key parameters of the system architecture and configurations; (3) employ suitable machine learning algorithms to improve model accuracy. We train models with five popular regression algorithms and conduct experiments on two distinct production HPC platforms. We find that the lasso and random forest models predict output performance with high accuracy on both of the target systems. We also explore use of the models to guide adaptation in I/O middleware systems, and show potential for improvements of at least 15% from model-guided adaptation on 70% of samples, and improvements up to 10 x on some samples for both of the target systems.