Search

Recent advances in artificial intelligence (AI) are radically changing how systems and applications are designed and developed. In this context, new requirements and regulations emerge, such as the AI Act, placing increasing focus on strict non-functional requirements, such as privacy and robustness, and how they are verified. Certification is considered the most suitable solution for non-functional verification of modern distributed systems, and is increasingly pushed forward in the verification of AI-based applications. In this paper, we present a novel dynamic malware detector driven by the requirements in the AI Act, which goes beyond standard support for high accuracy, and also considers privacy and robustness. Privacy aims to limit the need of malware detectors to examine the entire system in depth requiring administrator-level permissions; robustness refers to the ability to cope with malware mounting evasion attacks to escape detection. We then propose a certification scheme to evaluate non-functional properties of malware detectors, which is used to comparatively evaluate our malware detector and two representative deep-learning solutions in literature.

Federated learning is a distributed machine-learning technique that enables multiple devices to learn a shared model while keeping their local data private. The approach poses security challenges, such as model integrity, that must be addressed to ensure the reliability of the learned models. In this context, software-defined networking (SDN) can play a crucial role in improving the security of federated learning systems; indeed, it can provide centralized control and management of network resources, enforcement of security policies, and detection and mitigation of network-level threats. The integration of SDN with federated learning can help achieve a secure and efficient distributed learning environment. In this paper, an architecture is proposed to detect attacks on Federated Learning using SDN; furthermore, the machine learning model is deployed on a number of devices for training. The simulation results are carried out using the N-BaIoT dataset and training models such as Random Forest achieves 99.6%, Decision Tree achieves 99.8%, and K-Nearest Neighbor achieves 99.3% with 20 features.

Modern malware detection tools rely on special permissions to collect data that can reveal the presence of suspicious software within a machine. Typical data that they collect for this task are the set of system calls, the content of network traffic, file system changes, and API calls. However, giving access to these data to an externally created program means granting the company that created that software complete control over the host machine. This is undesirable for many reasons. In this work, we propose an alternative approach for this task, which relies on easily accessible data, information about system performances (CPU, RAM, disk, and network usage), and does not need high-level permissions to be collected. To investigate the effectiveness of this approach, we collected these data in the form of a multivalued time series and ran a number of malware programs in a suitably devised sandbox. Then – to address the fact that deep learning models need large training sets – we augmented the dataset using a deep learning generative model (a Generative Adversarial Network). Finally, we trained an LSTM (Long Short Term Memory) network to capture the malware behavioral patterns. Our investigation found that this approach, based on easy-to-collect information, is very effective (we achieved 0.99 accuracy), despite the fact that the data used for training the detector are substantially different from the ones specifically targeted for this purpose. The real and synthetic datasets, as well as corresponding source code, are publicly available.

Transfer Learning (TL) encompasses a number of Machine Learning Techniques that take a pre-trained model aimed at solving a task in a Source Domain and try to reuse it to improve the performance of a related task in a Target Domain An important issue in TL is that the effectiveness of those techniques is strongly dataset-dependent. In this work, we investigate the possible structural causes of the varying performance of Heterogeneous Transfer Learning (HTL) across domains characterized by different, but overlapping feature sets (this naturally determine a partition of the features into Source Domain specific subset, Target Domain specific subset, and shared subset). To this purpose, we use the Partial Information Decomposition (PID) framework, which breaks down the multivariate information that input variables hold about an output variable into three kinds of components: Unique, Synergistic, and Redundant. We consider that each domain can hold the PID components in implicit form: this restricts the information directly accessible to each domain. Based on the relative PID structure of the above mentioned feature subsets, the framework is able to tell, in principle: 1) which kind of information components are lost in passing from one domain to the other, 2) which kind of information components are at least implicitly available to a domain, and 3) what kind information components could be recovered through the bridge of the shared features. We show an example of a bridging scenario based on synthetic data.

This work surveys the research contributions of the last decade to the prediction of customer churn and adds a perspective toward what is yet to be reached. The main objective of this article is to report on (1) the methods and algorithms studied, the evaluation metrics adopted, and the results achieved, (2) the data used, and (3) the issues and limitations identified. Furthermore, the work highlights the gaps in the current literature and suggests a direction for future research.

Transfer Learning (TL) encompasses a number of Machine Learning Techniques that take a pre-trained model aimed at solving a task in a Source Domain and try to reuse it to improve the performance of a related task in a Target Domain An important issue in TL is that the effectiveness of those techniques is strongly dataset-dependent. In this work, we investigate the possible structural causes of the varying performance of Heterogeneous Transfer Learning (HTL) across domains characterized by different, but overlapping feature sets (this naturally determine a partition of the features into Source Domain specific subset, Target Domain specific subset, and shared subset). To this purpose, we use the Partial Information Decomposition (PID) framework, which breaks down the multivariate information that input variables hold about an output variable into three kinds of components: Unique, Synergistic, and Redundant. We consider that each domain can hold the PID components in implicit form: this restricts the information directly accessible to each domain. Based on the relative PID structure of the above mentioned feature subsets, the framework is able to tell, in principle: 1) which kind of information components are lost in passing from one domain to the other, 2) which kind of information components are at least implicitly available to a domain, and 3) what kind information components could be recovered through the bridge of the shared features. We show an example of a bridging scenario based on synthetic data.

Most data nowadays are stored in the cloud; therefore, cloud computing and its extension—fog computing—are the most in-demand services at the present time. Cloud and fog computing platforms are largely used by Internet of Things (IoT) applications where various mobile devices, end users, PCs, and smart objects are connected to each other via the internet. IoT applications are common in several application areas, such as healthcare, smart cities, industries, logistics, agriculture, and many more. Due to this, there is an increasing need for new security and privacy techniques, with attribute-based encryption (ABE) being the most effective among them. ABE provides fine-grained access control, enables secure storage of data on unreliable storage, and is flexible enough to be used in different systems. In this paper, we survey ABE schemes, their features, methodologies, benefits/drawbacks, attacks on ABE, and how ABE can be used with IoT and its applications. This survey reviews ABE models suitable for IoT platforms, taking into account the desired features and characteristics. We also discuss various performance indicators used for ABE and how they affect efficiency. Furthermore, some selected schemes are analyzed through simulation to compare their efficiency in terms of different performance indicators. As a result, we find that some schemes simultaneously perform well in one or two performance indicators, whereas none shines in all of them at once. The work will help researchers identify the characteristics of different ABE schemes quickly and recognize whether they are suitable for specific IoT applications. Future work that may be helpful for ABE is also discussed.

An essential component of the Internet of Things (IoT) is wireless sensor networks (WSNs). Since individual sensor nodes are strongly power-constrained, several techniques are adopted to save power. By grouping nodes into clusters—thus reducing the transmission distance between sensor nodes and the base station (BS)—a clustering protocol can ensure energy preservation and increase the lifetime of the network. However, current clustering techniques have problems with the clustering structure that negatively impact their performance. Whenever routing protocols were implemented for a longer period of time, it was observed that they had a higher rate of energy consumption, a shorter period of stability, and fewer data transfers to the BS. In this paper, an improved region-based routing protocol (REERP) is developed for wireless sensor networks in the IoT is developed. It is based on (i) the addition of new nodes to the already formed clusters, (ii) the selection of the new head node based on the amount of residual energy, (iii) the setup of the multi-hop communication in all the regions of network, and (iv) the utilization of the energy hole reduction method. All of these tactics increase the useful life of the network. Performance has been evaluated against (1) a stable election protocol, (2) a gateway energy-aware routing protocol, and (3) a heterogeneous gateway energy-aware routing protocol, and using the metrics lifetime, energy consumption, number of dead nodes, and number of packets sent to the base station vs. number of packets acquired by the base station. The results of the proposed routing protocol have been found to outperform the state-of-the-art approaches considered.

Several industries use wireless sensor networks (WSN) for various tasks such as monitoring, data transmission, and data gathering. They find applications in the industrial internet of things (IIoT). WSNs are utilized to track and monitor changes in the environment. Since they include multiple small sensor nodes (SN), they are severely constrained, so resource management geared toward energy efficiency is crucial in this kind of network. Minimizing the power to interpret, transmit, and store data between various sensors poses important challenges. Experts have considered various ways to address these issues that unavoidably affect the network’s performance: reducing energy usage while maintaining system throughput remains the primary research issue. Another important concern relates to network security. Specifically, intrusion detection and avoidance are major concerns. In this work, we introduce the meta-heuristic-based secure and energy-efficient routing (MHSEER) protocol for WSN-IIoT. The protocol learns the forwarding decisions using the number of hops, connection integrity characteristics, and accumulated remaining energy. To make the method more secure, the protocol also employs counter-encryption mode (CEM) to encrypt the data. A meta-heuristics study designed to achieve reliable learning is used in the suggested protocol. The protocol consists of two stages. The first stage uses a heuristics method to improve the option for dependable data routing. Security based on a computationally simple and random CEM is accomplished in the second stage. The proposed MHSEER protocol has been compared to the secure trust routing protocol for low power (Sectrust-RPL), heuristic-based energy-efficient routing (HBEER), secure and energy-aware heuristic-based routing (SEHR), and secure energy-aware meta-heuristic routing (SEAMHR) in terms of packet drop ratio, throughput, network delay, energy usage, and faulty pathways. The proposed protocol increases throughput to 95.81% and

The use of telerobots for surgery, military missions or rescue activities are increasingly commonplace, and so are guidelines to achieve an ergonomic design and optimal system-operator performance. However, in medical device manufacturing, in particular for fine-manipulation and highly precise operations, the existing human-machine interface and interaction (HMII) design standards to date are insufficient to ensure a cost-effective ergonomic teleoperation solution, that complies with the learnability, usability, dependability and efficiency required for this case-study. We analyze the most relevant human-system interface and interaction requirements for telerobotic systems applied to medical device manufacturing, and expand on the current standards with knowledge from recent research results in this field. We further our contribution by proposing a use-case-based telerobotic system architecture and experimental setup for human-centered evaluation of the telerobotic system HMII design.

The Retinex algorithms find wide applications as image enhancers, for their capability of preserving edges, while at the same time attenuating smooth gradients and chromatic dominants. They are characterized by the fact that the output chromatic intensity of a pixel is not determined in isolation (or looking only at the contiguous pixels) but through an operation of comparison to different local and remote areas of the image. This local/global comparison implies also a high computational cost for the algorithms: their complexity is not linear with the number of pixels; furthermore, the more systematic the comparison, the higher the complexity. Thus, most Retinex algorithms are unfit for real-time processing. The recent development of efficient Machine Learning architectures for Image Processing has raised the question of whether one of the Retinex "transforms"could be efficiently learned by training a feed-forward Artificial Neural Network, thus creating a model characterized by short processing time. Selecting a variant of the Random Spray Retinex model - FuzzyRSR - as representative of the Retinex family, and choosing suitably structured autoencoder neural networks, we found that we could accurately reproduce the Retinex effects. The computational cost of the training phase was moderate, while that of the inference phases was linear in the number of pixels, and three orders of magnitude lower than the one of FuzzyRSR, thus making the ANN implementation of Retinex suitable for real-time processing.

The Retinex theory, originally developed by Land and McCann as a computation model of the human color sensation, has become, with time, a pillar of digital image enhancement. In this area, the Retinex algorithm is widely used to improve the quality of any input image by increasing the visibility of its content and details, enhancing its colorfulness, and weakening, or even removing, some undesired effects of the illumination. The algorithm was originally described by its creators in terms of a sequence of image processing operations and was not fully formalized mathematically. Later, works focusing on aspects of the original formulation and adopting some of its principles tried to frame the algorithm within a mathematical formalism: this yielded every time a partial rendering of the model and resulted in several interesting model variants. The purpose of the present work is to fill a gap in the Retinex-related literature by providing a complete mathematical formalization of the original Retinex algorithm. The overarching goals of this work are to provide mathematical insights into the Retinex theory, promote awareness of the use of the model within image enhancement, and enable better appreciation of differences and similarities with later models based on Retinex principles. For this purpose, we compare our model with others proposed in the literature, paying particular attention to the work published in 2005 by Provenzi and others.

Random spray retinex (RSR) is an effective image enhancement algorithm owing to its effectiveness in improving the image quality. However, the computing complexity of the algorithm, the required hardware resources, and memory access hamper its deployment in many application scenarios, for instance, in IoT systems with limited hardware resources. With the rise of artificial intelligence (AI), the use of image enhancement has become essential for improving the performance of many emerging applications. In this paper, we propose the use of RSR as a preprocessing filter before the task of semantic segmentation of low-quality urban road scenes. Using the publicly available Cityscapes dataset, we compared the performance of a pre-trained deep semantic segmentation network on dark and noisy images with that of RSR preprocessed images. Our findings confirm the effectiveness of RSR in improving segmentation accuracy. In addition, to address the computational complexity and suitability of edge devices, we propose a novel and efficient implementation of RSR using resistive random access memory (RRAM) technology. This architecture provides highly parallel analog in-memory computing (IMC) capabilities. A detailed, efficient, and low-latency implementation of RSR using RRAM-CMOS technology is described. The design was verified using SPICE simulations with measured data from the fabricated RRAM and 65 nm CMOS technologies. The approach presented here represents an important step towards a low-complexity, real-time hardware-friendly architecture and the design of retinex algorithms for edge devices.

Spatial color algorithms (SCAs) are computer vision procedures widely used for image enhancement and human vision modeling. The main characteristic of SCA family is that they mimic the behavior of the human vision system (HVS), achieving in this way robustness and the capability to adjust their effect according to the image content. Here, we review 35 different, popular Retinex-inspired SCAs discussing and providing a set of measures for their evaluation in terms of image quality. To this purpose, we also introduce SCA-30, a real-world color image dataset made publicly available. The algorithms considered here include and spread from well-known Retinex implementations, Retinex variants, Milano-Retinex and related inspired enhancers, illumination/decomposition approaches, and deep learning-based techniques. Data and code used for the evaluation are made freely available to the community, to pursue further analysis and comparisons.

The Internet of Things (IoT) has recently emerged as a new family of technologies that allow a great number of things to be connected over heterogeneous networks. Conventional networks, however, face a technological barrier in efficiently handling such a large number of devices. The approach based on software-defined networks (SDNs), with its speed and flexibility, has recently been introduced into IoT area to potentially achieve scalability and adaptability, resulting in the SDN-IoT, a unique IoT design. In this work, we describe a new multiple-path routing model for the SDN-IoT. This model consists of two components: 1) a congested source path discovery technique; 2) a multi-path selection method taking into account route similarity and priority. In this paper we first describe the SDN based Quality of Service and highlight the unique features of this routing technique, then report on the performance evaluation of such system, which achieves an approximate 17% decrease in packet loss ratio over the existing approaches.

Recently, due to the rapid development of deep learning methods, there has been a growing interest in Neuro-symbolic Artificial Intelligence, which takes advantage of both explicit symbolic knowledge and statistical sub-symbolic neural knowledge representations. In sensor-based human performance prediction (HPP) for safety-critical applications, where maintaining optimal human and system performance is a major concern, neuro-symbolic AI systems can improve sensor-based HPP tasks in complex working settings. In this paper, we focus on the advantages of hybrid neuro-symbolic AI systems, present the outstanding challenges and propose possible solutions for HPP in the safety-critical application domain.

Given a class label y assigned by a classifier to a point x in feature space, the counterfactual generation task, in its simplest form, consists of finding the minimal edit that moves the feature vector to a new point x′, which the classifier maps to a pre-specified target class y′≠ y. Counterfactuals provide a local explanation to a classifier model, by answering the questions “Why did the model choose y instead of y′ : what changes to x would make the difference?". An important aspect in classification is ambiguity: typically, the description of an instance is compatible with more than one class. When ambiguity is too high, a suitably designed classifier can map an instance x to a class set Y of alternatives, rather than to a single class, so as to reduce the likelihood of wrong decisions. In this context, known as set-based classification, one can discuss set-based counterfactuals. In this work, we extend the counterfactual generation problem – normally expressed as a constrained optimization problem – to set-based counterfactuals. Using non-singleton counterfactuals, rather than singletons, makes the problem richer under several aspects, related to the fact that non-singleton sets allow for a wider spectrum of relationships among them: (1) the specification of the target set-based class Y′ is more varied (2) the target solution x′ that ought to be mapped to Y′ is not granted to exist, and, in that case, (3) since one might end up with the availability of a number of feasible alternatives to Y′, one has to include the degree of partial fulfillment of the solution into the loss function of the optimization problem.

An ISO 9001 audit can be seen as an independent risk assessment on the business, where each ‘Nonconformity” or “Opportunity For Improvement” is considered as a potential risk. Nevertheless, their actual impact on the business remains difficult to determine; as a consequence, the urgency of a mitigation plan at corporate level can sometimes be underestimated. This paper proposes a semi-quantitative risk assessment methodology on the ISO 9001 findings relying on a selected panel of experts. The experts’ responses are analyzed and validated using a specific statistics test for inter-rater reliability. The proposed methodology has been applied on real findings coming from ISO 9001 internal audits, involving 10 subject matter experts of 7 different countries.

This paper proposes a novel scheme to enhance the accuracy of packet-switched network synchronization systems by estimating path asymmetry (PA) using convolutional denoising autoencoders (CDAEs). Network synchronization is a key enabler of several emerging applications, with increasingly tight accuracy requirements especially for 5G. Path asymmetry, which arises due to physical and stochastic network conditions, severely degrades synchronization accuracy. In this paper, we propose a novel technique based on the IEEE Precision Time Protocol (PTP), which accurately reconstructs PA information from PTP packets. The proposed PA estimator can be integrated with existing synchronization systems as a pre-processing method to enhance the overall performance. Simulation results using industry-standard traffic profiles demonstrate significant improvements in PA estimation accuracy compared to the state of the art.

Given a class label y assigned by a classifier to a point x in feature space, the counterfactual generation task, in its simplest form, consists of finding the minimal edit that moves the feature vector to a new point x′, which the classifier maps to a pre-specified target class y′≠ y. Counterfactuals provide a local explanation to a classifier model, by answering the questions “Why did the model choose y instead of y′ : what changes to x would make the difference?". An important aspect in classification is ambiguity: typically, the description of an instance is compatible with more than one class. When ambiguity is too high, a suitably designed classifier can map an instance x to a class set Y of alternatives, rather than to a single class, so as to reduce the likelihood of wrong decisions. In this context, known as set-based classification, one can discuss set-based counterfactuals. In this work, we extend the counterfactual generation problem – normally expressed as a constrained optimization problem – to set-based counterfactuals. Using non-singleton counterfactuals, rather than singletons, makes the problem richer under several aspects, related to the fact that non-singleton sets allow for a wider spectrum of relationships among them: (1) the specification of the target set-based class Y′ is more varied (2) the target solution x′ that ought to be mapped to Y′ is not granted to exist, and, in that case, (3) since one might end up with the availability of a number of feasible alternatives to Y′, one has to include the degree of partial fulfillment of the solution into the loss function of the optimization problem.

This paper proposes a Deep Learning approach to radio signal de-noising. This approach is data-driven, thus it allows de-noising signals, corresponding to distinct protocols, without requiring explicit use of expert knowledge, in this way granting higher flexibility. The core component of the Artificial Neural Network architecture used in this work is a Convolutional De-noising AutoEncoder. We report about the performance of the system in spectrogram-based denoising of the protocol preamble across protocols of the IEEE 802.11 family, studied using simulation data. This approach can be used within a machine learning pipeline: the denoised data can be fed to a protocol classifier. A further perspective advantage of using the AutoEncoders in such a pipeline is that they can be co-trained with the downstream classifier (protocol detector), to optimize its accuracy.

Healthcare organizations and Health Monitoring Systems generate large volumes of complex data, which offer the opportunity for innovative investigations in medical decision making. In this paper, we propose a beetle swarm optimization and adaptive neuro-fuzzy inference system (BSO-ANFIS) model for heart disease and multi-disease diagnosis. The main components of our analytics pipeline are the modified crow search algorithm, used for feature extraction, and an ANFIS classification model whose parameters are optimized by means of a BSO algorithm. The accuracy achieved in heart disease detection is 99.1 % with 99.37 % precision. In multi-disease classification, the accuracy achieved is 96.08 % with 98.63 % precision. The results from both tasks prove the comparative advantage of the proposed BSO-ANFIS algorithm over the competitor models.

In this paper, we present a smart connected parking lots solution to automatically count and notify drivers about empty parking spots in major cities. As its name implies, the proposed smart IoT system has two operating phases: (i) continuous counting of empty spots in the monitored far-apart parking lots, and (ii) instantaneous driver notification through a lightweight MQTT mechanism. This notification system relies only on information collected from the pre-installed multimedia devices (no other apparatus installation or maintenance such as ground sensors is required). To validate the proper operation of our solution, we have implemented a small-scale version of it and assessed its performance while considering different classical and lightweight deep learning techniques (MobileNetV2, ResNet-50, YOLOv3, SSD-MobileNetV2, Tiny-YOLO, SqueezeDet, and SqueezeDet pruned with l1-norm). The experiments have confirmed the proper operation, efficiency, ease of deployment, and ease of extension of our system. They also confirmed that lightweight deep learning solutions are more adequate for small-sized resource-constrained embedded systems. They are more efficient in terms of inference time, size, resource consumption, and yield an accuracy that is close to that of classical solutions.

This paper proposes a novel Deep Learning (DL)-based approach for classifying the radio-access technology (RAT) of wireless emitters. The approach improves computational efficiency and accuracy under harsh channel conditions with respect to existing approaches. Intelligent spectrum monitoring is a crucial enabler for emerging wireless access environments that supports sharing of (and dynamic access to) spectral resources between multiple RATs and user classes. Emitter classification enables monitoring the varying patterns of spectral occupancy across RATs, which is instrumental in optimizing spectral utilization and interference management and supporting efficient enforcement of access regulations. Existing emitter classification approaches successfully leverage convolutional neural networks (CNNs) to recognize RAT visual features in spectrograms and other time-frequency representations; however, the corresponding classification accuracy degrades severely under harsh propagation conditions, and the computational cost of CNNs may limit their adoption in resource-constrained network edge scenarios. In this work, we propose a novel emitter classification solution consisting of a Denoising Autoencoder (DAE), which feeds a CNN classifier with lower di-mensionality, denoised representations of channel-corrupted spectrograms. We demonstrate—using a standard-compliant simulation of various RATs including LTE and four latest Wi-Fi standards—that in harsh channel conditions including non-line-of-sight, large scale fading, and mobility-induced Doppler shifts, our proposed solution outperforms a wide range of standalone CNNs and other machine learning models while requiring significantly less computational resources. The maximum achieved accuracy of the emitter classifier is 100%, and the average accuracy is 91% across all the propagation conditions.

In the automatic credit card transaction classification there are two phases: in the Real-Time (RT) phase the system decides quickly, based on the bare transaction information, whether to authorize the transaction; in the subsequent Near-Real-Time (NRT) phase, the system enacts a slower ex-post evaluation, based on a larger information context. The classification rules in the NRT phase trigger alerts on suspicious transactions, which are transferred to human investigators for final assessment. The management criteria used to select the rules, to be kept operational in the NRT pool, are traditionally based mostly on the performance of individual rules, considered in isolation; this approach disregards the non-additivity of the rules (aggregating rules with high individual precision does not necessarily make a high-precision pool). In this work, we propose to apply, to the rule selection for the NRT phase, an approach which assigns a normalized score to the individual rule, quantifying the rule influence on the overall performance of the pool. As a score we propose to use a power-index developed within Coalitional Game Theory, the Shapley Value (SV), summarizing the performance in collaboration. Such score has two main applications: (1) it can be used, within the periodic rule assessment process, to support the decision of whether to keep or drop the rule from the pool; (2) it can be used to select the k top-ranked rules, so as to work with a more compact rule set. Using real-world credit card fraud data containing approximately 300 rules and 3×105 transactions records, we show that: (1) this score fares better – in granting the performance of the pool – than the one assessing the rules in isolation; (2) that the same performance of the whole pool can be achieved keeping only one tenth of the rules — the top-k SV-ranked rules. We observe that the latter application can be reframed in terms of Feature Selection (FS) task for a classifier: we show that our approach is c

In order to keep in good long-term relationships with their main customers, Airline Cargo companies do not impose any fee for last minute cancellations of shipments. As a result, customers can book the same shipment on several cargo companies. Cargo companies try to balance cancellations by a corresponding volume of overbooking. However, the considerable uncertainty in the number of cancellations does not allow to fine-tune the optimal overbooking level, causing losses. In this work, we show how the deployment of cryptographic techniques, enabling the computation on private information of customers and companies data can improve the overall service chain, allowing for striking and enforcing better agreements. We propose a query system based on proxy re-encryption and show how the relevant information can be extracted, still preserving the privacy of customers’ data. Furthermore, we provide a Game Theoretic model of the use case scenario and show that it allows a more accurate estimate of the cancellation rates. This supports the reduction of the uncertainty and allows to better tune the overbooking level.

Anisotropic-diffusion is a commonly used signal preprocessing technique that allows extracting meaningful local characteristics from a signal, such as edges in an image and can be used to support higher-level processing tasks, such as shape detection. This paper presents a fully scalable CMOS-RRAM architecture of an edge-aware-anisotropic filtering algorithm aimed at computer vision applications. The CMOS circuitry controls the scale-space image data to perform pseudo-parallel in-memory computing and nonlinear processing through RRAM crossbar. The arithmetic operations for in-memory computation of brightness gradients are efficiently accumulated to produce the enhanced image in several iterations. The proposed architecture uses single RRAM as a computing and storage element to perform both arithmetic operations and accumulations. Thanks to the in-memory computation, memory accesses and arithmetic operations are reduced by 64% and 92%, respectively, compared to traditional digital implementations. This, in turn, results in a potential reduction of power and area costs of about 75% and 85%, respectively. The processing time is also reduced by 67%.

The opening of the unlicensed radio spectrum creates new opportunities and new challenges for communication technology that can be faced by Machine Learning techniques. In this work, we discuss the potential benefits and the challenges with reference to the recent research developments in this area. Applications go from channel estimation to Signal quality control, and from signal classification to action control. We survey Machine learning and Deep Learning algorithms with possible radio applications and highlight the corresponding challenges.

Within the paradigm of Mobile Cloud Computing (MCC) mobile devices such as mobile phones and tablets can unload computation to a local Cloud consisting of both static and mobile devices. Due to the proximity, such Cloud can better fulfill service and latency requirements for QoS sensitive applications. However, in order to function properly, such mechanics requires a collaborative approach that not all the devices might follow: some can contribute much less than others and deviate from the Service Level Agreement. This creates a free-riding problem in MCC and a corresponding QoS issue. An approach towards this problem consists of providing incentives to nodes so that they act as inspectors and occasionally audit the recent behavior of the nodes with which they interact. This double role of potential inspectee and inspectors can be modeled within Game Theory (GT) to predict the behavior of the agents. In this work, we show that an appropriate model for this game is a symmetric four-strategies social dilemma. The solution (a.k.a. equilibrium) of the game provides a unique way to set the incentives so as to drive the system toward the desired behavior. However, even the hybrid inspector-inspectee agent approach is in principle open to a potential flaw: the possibility of inspector-inspectee collusion. The main contribution of this work is that the collusion behavior is not an equilibrium of the game, thus hybrid agent rational players should never collude.

In this work we address the problem of critical source selection in social sensing. We propose an approach to the ranking of information streams, which is aware of the interdependence among streams (redundancy and synergies), of the cost of individual streams, and of the cost related to the integration of multiple streams. The method is based on the use of the Coalitional Game Theory concept of Power Index, and relies on the polynomial-time estimate of the stream sets characteristics. With respect to other works using a power index, the method takes into account that the problem has a non-trivial cost structure.

Within the paradigm of Mobile Cloud Computing (MCC) mobile devices such as mobile phones and tablets can unload computation to a local Cloud consisting of both static and mobile devices. Due to the proximity, such Cloud can better fulfill service and latency requirements for QoS sensitive applications. However, in order to function properly, such mechanics requires a collaborative approach that not all the devices might follow: some can contribute much less than others and deviate from the Service Level Agreement. This creates a free-riding problem in MCC and a corresponding QoS issue. An approach towards this problem consists of providing incentives to nodes so that they act as inspectors and occasionally audit the recent behavior of the nodes with which they interact. This double role of potential inspectee and inspectors can be modeled within Game Theory (GT) to predict the behavior of the agents. In this work, we show that an appropriate model for this game is a symmetric four-strategies social dilemma. The solution (a.k.a. equilibrium) of the game provides a unique way to set the incentives so as to drive the system toward the desired behavior. However, even the hybrid inspector-inspectee agent approach is in principle open to a potential flaw: the possibility of inspector-inspectee collusion. The main contribution of this work is that the collusion behavior is not an equilibrium of the game, thus hybrid agent rational players should never collude.

In this work we address the problem of critical source selection in social sensing. We propose an approach to the ranking of information streams, which is aware of the interdependence among streams (redundancy and synergies), of the cost of individual streams, and of the cost related to the integration of multiple streams. The method is based on the use of the Coalitional Game Theory concept of Power Index, and relies on the polynomial-time estimate of the stream sets characteristics. With respect to other works using a power index, the method takes into account that the problem has a non-trivial cost structure.

Traditional authentication systems use alphanumeric or graphical passwords, or token-based techniques that require 'something you know and something you have'. The disadvantages of these systems include the risks of forgetfulness, loss, and theft. To address these shortcomings, biometric authentication is rapidly replacing traditional authentication methods and is becoming a part of everyday life. The electrocardiogram (ECG) is one of the most recent traits considered for biometric purposes. In this work we describe an ECG-based authentication system suitable for security checks and hospital environments. The proposed system will help investigators studying ECG-based biometric authentication techniques to define dataset boundaries and to acquire high-quality training data. We evaluated the performance of the proposed system and found that it could achieve up to the 92% identification accuracy. In addition, by applying the Amang ECG (amgecg) toolbox within MATLAB, we investigated the two parameters that directly affect the accuracy of authentication: The ECG slicing time (sliding window) and the sampling time period, and found their optimal values

We investigate the hypothesis that long-term memory in populations of agents can lead to counterproductive emergent properties at the system level. Our investigation is framed in the context of a discrete, one-dimensional road-traffic congestion model: we investigate the influence of simple cognition in a population of rational commuter agents that use memory to optimise their departure time, taking into account congestion delays on previous trips. Our results differ from the well-known minority game in that crowded slots do not carry any explicit penalty. We use Markov chain analysis to uncover fundamental properties of this model and then use the gained insight as a benchmark. Then, using Monte Carlo simulations, we study two scenarios: one in which “myopic” agents only remember the outcome (delay) of their latest commute, and one in which their memory is practically infinite. We show that there exists a trade-off, whereby myopic memory reduces congestion but increases uncertainty, while infinite memory does the opposite. We evaluate the performance against the optimal distribution of departure times (i.e. where both delay and uncertainty are minimised simultaneously). This optimal but unstable distribution is identified using a genetic algorithm.

Glare is always present in optical acquisition systems, such as photocameras or the eye-bulb. As a consequence, images captured by sensors do not represent an accurate reproduction of the scene, but rather a combination of scene content and glare. We discuss the reasons why this unwanted addition of spread light cannot be removed from an acquired image. To this aim, we cast the problem of glare-removal into an estimation task and focus on the aspects that make the unfolding of glare an ill-posed or ill-conditioned problem-such as nonlinearity, information loss, or eye model uncertainty. For each mechanism of glare formation, we point to the corresponding influence in terms of ill-posedness and ill-conditioning of the problem. We do not aim at proposing or reviewing solutions to the glare problem but rather at identifying more precisely the challenges it poses.

A challenge in designing cooperative distributed systems is to develop feasible and cost-effective mechanisms to foster cooperation among selfish nodes, i.e., nodes that strategically deviate from the intended specification to increase their individual utility. Finding a satisfactory solution to this challenge may be complicated by the intrinsic characteristics of each system, as well as by the particular objectives set by the system designer. Our previous work addressed this challenge by proposing RACOON, a general and semi-Automatic framework for designing selfishness-resilient cooperative systems. RACOON relies on classical game theory and a custom built simulator to predict the impact of a fixed set of selfish behaviours on the designer's objectives. In this paper, we present RACOON++, which extends the previous framework with a declarative model for defining the utility function and the static behaviour of selfish nodes, along with a new model for reasoning on the dynamic interactions of nodes, based on evolutionary game theory. We illustrate the benefits of using RACOON++ by designing three cooperative systems: A peer-To-peer live streaming system, a load balancing protocol, and an anonymous communication system. Extensive experimental results using the state-of-The-Art PeerSim simulator verify that the systems designed using RACOON++ achieve both selfishness-resilience and high performance.

We investigated the use of a Deep Learning approach to radio signal de-noising. This data-driven approach has does not require explicit use of expert knowledge to set up the parameters of the denoising procedure and grants great flexibility across many channel conditions. The core component used in this work is a Convolutional De-noising AutoEncoder, known to be very effective in image processing. The key of our approach consists in transforming the radio signal into a representation suitable to the CDAE: we transform the time-domain signal into a 2D signal using the Short Time Fourier Transform. We report about the performance of the approach in preamble denoising across protocols of the IEEE 802.11 family, studied using simulation data. This approach could be used within a machine learning pipeline: the denoised data can be fed to a protocol classifier. A perspective advantage of using the AutoEncoders in that pipeline is that they can be co-trained with the downstream classifier, to optimize the classification accuracy.

Real-life event logs are typically much less structured and more complex than the predefined business activities they refer to. Most of the existing process mining techniques assume that there is a one-to-one mapping between process model activities and events recorded during process execution. Unfortunately, event logs and process model activities are defined at different levels of granularity. The challenges posed by this discrepancy can be addressed by means of log-lifting. In this work we develop a machine-learning-based framework aimed at bridging the abstraction level gap between logs and process models. The proposed framework operates of two main phases: log segmentation and machine-learning-based classification. The purpose of the segmentation phase is to identify the potential segment separators in a flow of low-level events, in which each segment corresponds to an unknown high-level activity. For this, we propose a segmentation algorithm based on maximum likelihood with n-gram analysis. In the second phase, event segments are mapped into their corresponding high-level activities using a supervised machine learning technique. Several machine learning classification methods are explored including ANNs, SVMs, and random forest. We demonstrate the applicability of our framework using a real-life event log provided by the SAP company. The results obtained show that a machine learning approach based on the random forest algorithm outperforms the other methods with an accuracy of 96.4%. The testing time was found to be around 0.01s, which makes the algorithm a good candidate for real-time deployment scenarios.

Unstructured networks are characterized by constrained resources and require protocols that efficiently utilize bandwidth and battery power. Probabilistic flooding, allows nodes to rebroadcast RREQ packets with some probability p, thus reducing the overhead. The key issue in of this algorithm consists of determining p. The techniques proposed so far either use a fixed p determined by a priori considerations, or a p variable from one node to the other - set, for instance based on node degree or distance between source and destination - or even a dynamic p based on the number of redundant messages received by the nodes. In order to make the computation of forwarding probability p works optimally regardless of changing of topology, we propose to set p based on the node role within the message dissemination process. Specifically, we propose to identify such role based on the nodes' clustering coefficients (the lower the coefficient, the higher the forwarding probability). The performance of the algorithm is evaluated in terms of routing overhead, packet delivery ratio, and end-to-end delay. The algorithm pays a price in terms of computation time for discovering the clustering coefficient, however reduces unnecessary and redundant control messages and achieves a significant improvements in both dense and sparse networks in terms of packet delivery ratio. We compare by simulation the performance of this algorithm with the one of the most representative competing algorithms.

In contrast to single-view learning, multi-view learning trains simultaneously distinct algorithms on disjoint subsets of features (the views), and jointly optimizes them, so that they come to a consensus. Multi-view learning is typically used when the data are described by a large number of features. It aims at exploiting the different statistical properties of distinct views. A task to be performed before multi-view learning - in the case where the features have no natural groupings - is multi-view generation (MVG): it consists in partitioning the feature set in subsets (views) characterized by some desired properties. Given a dataset, in the form of a table with a large number of columns, the desired solution of the MVG problem is a partition of the columns that optimizes an objective function, encoding typical requirements. If the class labels are available, one wants to minimize the inter-view redundancy in target prediction and maximize consistency. If the class labels are not available, one wants simply to minimize inter-view redundancy (minimize the information each view has about the others). In this work, we approach the MVG problem in the latter, unsupervised, setting. Our approach is based on the transposition of the data table: the original instance rows are mapped into columns (the 'pseudo-features'), while the original feature columns become rows (the 'pseudo-instances'). The latter can then be partitioned by any suitable standard instance-partitioning algorithm: the resulting groups can be considered as groups of the original features, i.e. views, solution of the MVG problem. We demonstrate the approach using k-means and the standard benchmark MNIST dataset of handwritten digits.