• Performance guarantees for rule-based classification models. • Novel score function accounting for rule overlaps via geometrical rule similarity. • Conformal prediction-guided tuning of rules via conformal critical set. Computer scientists consider an artificial intelligence system safe and trustworthy if it fulfills four pillars: robustness, transparency, fairness, and privacy. We propose a fifth fundamental aspect: conformal guarantee, that is, the probabilistic assurance that the system will behave as expected. We introduce CONFIDERAI (Conformal Interpretable by Design Explainable and Reliable Artificial Intelligence), a new score function for binary rule-based classifiers depending on both rules’ performance and geometry; the latter includes both the position of points within rule boundaries and rule overlaps, these being quantified via geometrical rule similarity. Furthermore, we address the problem of individuating regions in the feature space in which conformal guarantees are satisfied, by defining the concept of conformal critical set (CCS). The overall method is tested with promising results, comparable in efficiency to traditional scores, on ten datasets of real-world interest, such as domain name server tunneling detection and cardiovascular disease prediction. Moreover, newly generated rules from CCS resulted into an improved precision and reduced error on a target class, thus avoiding prediction failures in safety-critical contexts.

Data augmentation represents an opportunity for Artificial Intelligence (AI) applications, as it aims at creating new synthetic data based on an existing baseline. In this paper, we present a new evaluation framework for Generative Adversarial Networks (GANs), a data augmentation technique, in multivariate data classification contexts. The goal is not limited to assessing the performance variations obtained through GANs, but also to inspect results with explainable AI (XAI) tools, understanding how GANs work and, finally, exploiting them to discover new knowledge. To this aim, we adopt the Logic Learning Machine (LLM) for performance assessment and rule extraction, and introduce a new measure of rule similarity to compare different artificial datasets. We apply the methodology on two case studies , activity recognition and physical fatigue detection, confirming that GANs can help in overcoming limitations of original datasets and lead to new discoveries.

ABSTRACT The global adoption of artificial intelligence (AI) has increased dramatically in recent years, becoming commonplace in many fields. Such a pervasiveness has led to changes in how AI is perceived, strengthening discussions on its societal consequences. Thus, a new class of requirements for AI‐based solutions emerged. Broadly speaking, those on “explainability” aim to provide a transparent representation of the (often opaque) reasoning method that an AI‐based solution uses when prompted. This work presents a methodology for validating a class of explainable AI (XAI) models, called deterministic rule‐based models, which are used for expressing an explainable approximation of classifiers based on machine learning. The validation methodology combines logical deduction with constraint‐based reasoning in numerical domains, and it either succeeds or returns quantitative estimations of the invalid deviations found. This information allows us to assess the correctness of an XAI model, or in the case of deviations, to evaluate if it still can be deemed acceptable. The validation methodology has been applied to a simulation‐based study where the decision‐making process copes with the spread of SARS‐COV‐2 inside a railway station. The considered case study is a controlled but nontrivial example that shows the overall applicability of the methodology.

Recent advances in Artificial Intelligence (AI) in healthcare are driving research into solutions that can provide personalized guidance. For these solutions to be used as clinical decision support tools, the results provided must be interpretable and consistent with medical knowledge. To this end, this study explores the use of explainable AI to characterize the risk of developing cardiovascular disease in patients diagnosed with chronic obstructive pulmonary disease. A dataset of 9613 records from patients diagnosed with chronic obstructive pulmonary disease was classified into three categories of cardiovascular risk (low, moderate, and high), as estimated by the Framingham Risk Score. Counterfactual explanations were generated with two different methods, MUlti Counterfactuals via Halton sampling (MUCH) and Diverse Counterfactual Explanation (DiCE). An error control mechanism is introduced in the preliminary classification phase to reduce classification errors and obtain meaningful and representative explanations. Furthermore, the concept of counterfactual conformity is introduced as a new way to validate single counterfactual explanations in terms of their conformity, based on proximity with respect to the factual observation and plausibility. The results indicate that explanations generated with MUCH are generally more plausible (lower implausibility) and more distinguishable (higher discriminative power) from the original class than those generated with DiCE, whereas DiCE shows better availability, proximity and sparsity. Furthermore, filtering the counterfactual explanations by eliminating the non-conformal ones results in an additional improvement in quality. The results of this study suggest that combining counterfactual explanations generation with conformity evaluation is worth further validation and expert assessment to enable future development of support tools that provide personalized recommendations for reducing individual risk by targeting specific subsets of biomarkers.

The supervised classification recognizes patterns in the data to separate classes of behaviors. Canonical solutions contain misclassification errors that are intrinsic to the numerical approximating nature of machine learning (ML). The data analyst may minimize the classification error on a class at the expense of increasing the error of the other classes. The error control of such a design phase is often done in a heuristic manner. In this article, it is key to develop theoretical foundations capable of providing probabilistic certifications to the obtained classifiers. In this perspective, we introduce the concept of probabilistic safety region to describe a subset of the input space in which the number of misclassified instances is probabilistically controlled. The notion of adjustable classifiers, a special class of classifiers that share the property of being controllable by a scalar parameter, is then exploited to link the tuning of ML with error control. Several tests and examples corroborate the approach. They are provided through the synthetic data in order to highlight all the steps involved, as well as notable benchmark datasets and a smart mobility application.

Abstract Conformal predictions make it possible to define reliable and robust learning algorithms. But they are essentially a method for evaluating whether an algorithm is good enough to be used in practice. To define a reliable learning framework for classification from the very beginning of its design, the concept of scalable classifier was introduced to generalize the concept of classical classifier by linking it to statistical order theory and probabilistic learning theory. In this paper, we analyze the similarities between scalable classifiers and conformal predictions by introducing a new definition of a score function and defining a special set of input variables, the conformal safety set, which can identify patterns in the input space that satisfy the error coverage guarantee, i.e., that the probability of observing the wrong (possibly unsafe) label for points belonging to this set is bounded by a predefined $$\varepsilon$$ ε error level. We demonstrate the practical implications of this framework through an application in cybersecurity for identifying DNS tunneling attacks. Our work contributes to the development of probabilistically robust and reliable machine learning models.

The aim of this study is to apply and characterize eXplainable AI (XAI) to assess the quality of synthetic health data generated using a data augmentation algorithm. In this exploratory study, several synthetic datasets are generated using various configurations of a conditional Generative Adversarial Network (GAN) from a set of 156 observations related to adult hearing screening. A rule-based native XAI algorithm, the Logic Learning Machine, is used in combination with conventional utility metrics. The classification performance in different conditions is assessed: models trained and tested on synthetic data, models trained on synthetic data and tested on real data, and models trained on real data and tested on synthetic data. The rules extracted from real and synthetic data are then compared using a rule similarity metric. The results indicate that XAI may be used to assess the quality of synthetic data by (i) the analysis of classification performance and (ii) the analysis of the rules extracted on real and synthetic data (number, covering, structure, cut-off values, and similarity). These results suggest that XAI can be used in an original way to assess synthetic health data and extract knowledge about the mechanisms underlying the generated data.

Data augmentation is a widespread innovative technique in Artificial Intelligence: it aims at creating new synthetic data given an existing real baseline, thus allowing to overcome the issues arising from the lack of labelled data for proper training of classification algorithms. Our paper focuses on how a common data augmentation methodology, the Generative Adversarial Networks (GANs), which is widespread for images and timeseries data, can be also applied to generate multivariate data. We propose a novel scheme for GANs evaluation, based on the performance of an explainable AI (XAI) algorithm and an innovative definition of rule similarity. In particular, we will consider an application dealing with the augmentation of Inertial Movement Units (IMU) data for physical fatigue monitoring in two age subgroups (under and over 40 years old) of the original data. We will show how our innovative rule similarity metric can drive the selection of the best fake dataset among a set of different candidates, corresponding to different GAN training runs.

One of the current gaps in teleaudiology is the lack of methods for adult hearing screening viable for use in individuals of unknown language and in varying environments. We have developed a novel automated speech-in-noise test that uses stimuli viable for use in non-native listeners. The test reliability has been demonstrated in laboratory settings and in uncontrolled environmental noise settings in previous studies. The aim of this study was: (i) to evaluate the ability of the test to identify hearing loss using multivariate logistic regression classifiers in a population of 148 unscreened adults and (ii) to evaluate the ear-level sound pressure levels generated by different earphones and headphones as a function of the test volume. The multivariate classifiers had sensitivity equal to 0.79 and specificity equal to 0.79 using both the full set of features extracted from the test as well as a subset of three features (speech recognition threshold, age, and number of correct responses). The analysis of the ear-level sound pressure levels showed substantial variability across transducer types and models, with earphones levels being up to 22 dB lower than those of headphones. Overall, these results suggest that the proposed approach might be viable for hearing screening in varying environments if an option to self-adjust the test volume is included and if headphones are used. Future research is needed to assess the viability of the test for screening at a distance, for example by addressing the influence of user interface, device, and settings, on a large sample of subjects with varying hearing loss.

Low-rate denial of service attacks are considered a serious threat for network systems. In this paper, we investigate such topic, by proposing a novel anomaly-based intrusion detection system. We validate the proposed system and report the weaknesses we have found. By working from the attacker's perspective, we also try to elude the proposed algorithm. Results show that in order to avoid detection, the attacker would require high-bandwidth to perpetrate the attack. The proposed method should therefore be considered an efficient method to detect running Slow DoS Attacks.

The objective of this work is the design of a technological platform for remote monitoring of patients with Chronic Obstructive Pulmonary Disease (COPD). The concept of the framework is a breakthrough in the state of medical, scientific and technological art, aimed at engaging patients in the treatment plan and supporting interaction with healthcare professionals. The proposed platform is able to support a new paradigm for the management of patients with COPD, by integrating clinical data and parameters monitored in daily life using Artificial Intelligence algorithms. Therefore, the doctor is provided with a dynamic picture of the disease and its impact on lifestyle and vice versa, and can thus plan more personalized diagnostics, therapeutics, and social interventions. This strategy allows for a more effective organization of access to outpatient care and therefore a reduction of emergencies and hospitalizations because exacerbations of the disease can be better prevented and monitored. Hence, it can result in improvements in patients' quality of life and lower costs for the healthcare system.

Offering very large data rates is one of the main objectives of Very High Throughput Satellite (VHTS) systems to boost enhanced Mobile Broadband (eMBB) services in converged 5G-satellite systems. To this end, the exploitation of Q/V frequency bands for the satellite feeder link is a key factor for guaranteeing unprecedented data rates provided that efficient handover algorithms are implemented to counteract link outage events caused by severe weather impairments. Moreover, Quality of Service (QoS) of data flows is typically affected when gateway handover is initiated, hence calling for sophisticated ground segment management solutions. In this light, this paper proposes a novel gateway handover strategy and validates its design through simulation campaigns, whose preliminary results show important performance gains with respect to other solutions available from the existing literature.

The Safe Cooperating Cyber-Physical Systems using Wireless Communication (SafeCop) project addresses safety-related issues in cooperating cyber-physical systems. These systems, characterised by wireless communications, multiple stakeholders, and variable operating environments, are called Cooperative Open Cyber-Physical Systems (CO-CPS). CO-CPSs can successfully address several societal challenges -- cooperative vehicles have been shown to reduce fuel consumption as well as the number of accidents. A vehicle-to-infrastructure (V2I) cooperation for the traffic management scenario is therefore considered as a key use cases of SafeCop. In this paper, we outline the V2I traffic management scenario, assess the research goals that arise from it, and provide an overview of the architecture of the demonstrator, as well as a roadmap for its development and evaluation.

The design of High Throughput Satellite (HTS) systems builds on the concept of Smart Gateway Diversity (SGD) to exploit the spatial diversity of gateways in case of feeder link outage, occurring because of atmospheric impairments introduced in Extremely High Frequency (EHF) frequency bands. The gateway handover procedure requires precise prediction algorithms and coordination among different network elements. This paper presents novel outage prediction algorithms based on machine learning concepts, integrated in the framework of SDN architectures, to efficiently orchestrate the gateway handover operations. Simulation campaigns prove the validity of the proposed concept and shed light on the potentials of the SDN architecture in future HTS systems.

Due to their recent appearance and the reduced requirements in terms of network bandwidth, Slow Denial of Service Attacks detection represents a particularly challenging problem. This paper presents a novel detection method, analyzing spectral features of the network traffic over small time horizons. The proposed method has been validated by extrapolating data referred to real traffic traces, elaborated over the Local Area Network of our research institute. We have considered different kinds of attacks and results show how the proposed approach is reliable and applicable also in other cybersecurity contexts.

An estimation setting is considered, where a number of sensors transmit their observations of a physical phenomenon, described by one or more random variables, to a sink over noisy communication channels. The goal is to minimize a quadratic distortion measure (Minimum Mean Square Error - MMSE) under a global power constraint on the sensors' transmissions. Linear MMSE encoders and decoders, parametrically optimized in encoders' gains, Shannon-Kotel'nikov mappings, and nonlinear parametric functional approximators (neural networks) are investigated and numerically compared, highlighting subtle differences in sensitivity and achievable performance.

The paper deals with DNS tunneling detection by means of simple supervised learning schemes, applied to statistical features of DNS queries and answers. DNS traffic samples are used by exploiting the content of the entire DNS database, thus avoiding socket-by-socket inspection. Specific attention is devoted to the detection of small portions of malicious data, hidden by regular DNS communication. Second and third level DNS domains are analyzed. Despite the simplicity of the mechanism, good results are obtained by replicating individual detections over successive samples over time and by making a global decision through a majority voting scheme. In this perspective, an empirical trade-off is found between fast and reliable detections.

The authors consider the acquisition of measurements from a source, representing a physical phenomenon, by means of sensors deployed at different distances, and measuring random variables (r.v.'s) that are correlated with the source output. The acquired values are transmitted over a wireless channel to a sink, where an estimation of the source has to be constructed, according to a given distortion criterion. In the presence of Gaussian random variables (r.v.'s) and a Gaussian vector channel, the authors are seeking optimum real-time joint source-channel encoder–decoder pairs that achieve a distortion sufficiently close to the theoretically optimal one, under a global resource constraint, by activating only a subset of the sensors. The problem is posed in a team decision theoretic framework, and the optimal strategies are approximated by means of neural networks. The analysis investigates the generalisation capabilities of the proposed approach, by showing insights into the structure of the problem. The surprising outcome is that a quasi-static application of the approach reveals to be sufficient to maintain quasi-optimal performance under a dynamic environment (e.g. with respect to nodes' positions).

This paper shows a detailed performance evaluation of the bandwidth allocation algorithm presented in [1] by the same research group and briefly reminded here. Telecommunications networks are composed of functional layers acting in cascade. The mentioned algorithm adapts the bandwidth to be allocated to a buffer which conveys heterogeneous traffic in a layer-in-cascade functional model, is based only on measures and does not use closed-form expressions, a-priori information about traffic statistical properties, and assumptions about buffer dimension. It is called RCBC (Reference Chaser Bandwidth Control) in [1]. The performance evaluation presented here is aimed at getting a general validation by testing RCBC over multiple DVB-based satellite scenarios and by analyzing the robustness of the algorithm against parameter variations. The performance evaluation is carried out through multiple performance metrics and, in consequence, the results are interpreted through MADM (Multi Attribute Decisions Making) theory, which is appropriate to take decisions in presence of multiple performance indicators.

The paper studies a protocol and a control architecture aimed at implementing bandwidth and call admission optimization over a DVB Return Channel Satellite Terminal (RCST) under Quality of Service (QoS) contraints. The approach can be applied in all cases where traffic flows, coming from a terrestrial portion of the network, are merged together within a single DVB flow, which is then forwarded over the satellite channel. The exact bandwidth to be provided to the flows entering the DVB layer is unknown because equivalent bandwidth techniques guarantee only approximate solutions in heterogeneous conditions. The paper firstly outlines merging operations of the flows over the DVB system and then it specifies the entities to be used at layer 3 and layer 2 of the RCST to coordinate the necessary actions to guarantee the QoS with minimum resource allocation.

The paper addresses resilience over Ethernet networks using the Rapid Spanning Tree Protocol (RSTP). The topic constitutes an open issue of debate as clear indications on the real RSTP performance can hardly be found from the literature. Actually, the complicated protocol structure makes the analysis intricate and unsuitable for generalization. Moreover, the presence of other resilience algorithms, whose mechanisms and rules are explicitly designed for resilience, solves the problem beyond the application of RSTP. Even though those solutions are actually more efficient than RSTP, they are more expensive. In this perspective, the purposes of this paper are to critically evaluate the intrinsic limitations of RSTP and propose some simple protocol modifications to speed up its performance. The analysis validates the achievable performance as a trade-off between fast reactions and bandwidth overhead.

Telecommunications networks are composed of functional layers acting in cascade. Quality of service (QoS) derives from the action of each layer that must assure a specific level of quality to the upper layer in terms of performance parameters. The action is called vertical QoS mapping and is provided through algorithms that compute the bandwidth necessary so to assure the requested QoS when information is transferred from one layer to the next one below. This paper proposes a scheme that adapts the bandwidth to be allocated to a buffer which conveys heterogeneous traffic (both concerning traffic sources and QoS requirements) in a layer-in-cascade model. The proposed algorithm is based only on measures and does not use closed-form expressions, a-priori information about traffic statistical properties, and assumptions about buffer dimension. It is called WI-RCBC (wireless interface reference chaser bandwidth control).

We consider the acquisition of measurements from a source, representing a physical phenomenon, by means of sensors deployed at different distances, and measuring random variables that are correlated with the source output. The acquired values are transmitted to a sink, where an estimation of the source has to be constructed, according to a given distortion criterion. In the presence of Gaussian random variables and a Gaussian vector channel, we are seeking optimum real-time joint source-channel encoder-decoder pairs that achieve a distortion sufficiently close to the theoretically optimal one, under a global power constraint, by activating only a subset of the sensors. The problem is posed in a team decision theoretic framework, and the optimal strategies are approximated by means of neural networks. We compare the solution with the results obtained by heuristically choosing a subset of the sensors on the basis of successive simulations under a fixed topology.

The paper proposes a novel measurement-based equivalent bandwidth technique that computes the bandwidth to be allocated to a buffer which conveys heterogeneous traffic (both concerning traffic sources and QoS requirements), without using any closed-form expression. The effectiveness of the algorithm is checked through simulation analysis.

The paper deals with protocol architectures to support mapping of quality of service (QoS) between protocol layers of telecommunication networks. The Technology Independent - Service Access Point studied to this aim within the ETSI committee is summarized and analyzed. Inherent QoS mapping operations introduce the generalization of the regular concept of equivalent bandwidth (EqB). Some performance evaluation is proposed to highlight EqB dimensioning.

The problem of pricing for a telecommunication network is investigated with respect to the users' sensitivity to the pricing structure. A functional optimization problem is formulated, in order to compute price reallocations as functions of data collected in real time during the network evolution. No a-priori knowledge about the users' utility functions and the traffic demands is required, since adaptive reactions to the network conditions are sought in real time. To this aim, a neural approximation technique is studied to exploit an optimal pricing control law, able to counteract traffic changes with a small online computational effort.

In this paper, the IEEE 802.16 protocol is investigated with respect to the bandwidth provision problem arising at the medium access control (MAC) layer. The aim is to optimally tune the resource allocation to match QoS requirements. Traffic flows are originated at network layers overlying the 802.16 protocol stack. This leads to the investigation of a novel control algorithm, suited to optimal bandwidth allocation and call admission control in the presence of statistically heterogeneous flows. Specific implementation details are provided to match the application of the control algorithm using the regular 802.16 request-grant protocol. Simulation results validate the proposed approach.

An optimization problem for QoS mapping over protocol layers is formalized in this work by taking the ETSIsatelliteindependent-serviceaccesspoint(SI-SAP) as technological reference. The joint optimization of loss and delay performance metrics, together with the presence of SI-SAP QoS mapping operations, introduces the generalization of the regular concept ofequivalentbandwidth. This leads to the study of a proper measurement-based control, able to capture the most stringent QoS requirement over time (between loss and delay) and the related instantaneous bandwidth need at the lower layer of the protocol stack. The control methodology is tested through real fading and traffic traces to highlight its effectiveness for real time control of QoS mapping operations.

This paper deals with a NP-hard resource allocation problem for a satellite network. An approach based on the estimation of the gradient of a cost function, obtained through a "relaxed continuous extension" of the discrete constraint set, is proposed. Since neither closed forms of the performance measure, nor additional feedbacks on the statistical properties of the traffic sources are requested, the proposed approach reveals to be suitable for optimizing the resource allocation in real life case studies, where the application of specific certainty equivalent assumptions is impractical

The quality of service (QoS) provision requires the cooperation of all network layers from bottom-to-top. More specifically, in the ETSI broadband satellite multimedia architecture, the physical layers are isolated from the rest by a satellite independent-service access point (SI-SAP). The structure of SI-SAP protocol model "opens the door" to the problem of mapping the performance requests of satellite independent (SI) layers over (SD) satellite dependent technology. In this perspective, we investigate here a novel control algorithm for QoS mapping at the SI-SAP interface. Through the adoption of infinitesimal perturbation analysis, we are able to capture the "bandwidth need" of the different flows when they are conveyed to the SD core. Simulation results validate the proposed approach.

In this paper, a novel call admission control (CAC) problem is investigated in relation to the pricing structure of a telecommunication network, in which both guaranteed performance (GP) and best effort (BE) services are offered. The user's sensitivity to the prices is described through utility functions. An original decision making process is studied to decentralize the proposed CAC mechanism. To this aim, a neural approximation technique is investigated to exploit different decision makers, distributed in the network and performing the CAC decisions. Simulation results show how sub-optimal CAC decisions are obtained in a decentralized fashion and with a small on-line computational effort.

A resource allocation problem for a satellite network is considered, where variations of fading conditions are added to those of traffic load. Since the capacity of the system is finite and divided in finite discrete portions, the resource allocation problem reveals to be a discrete stochastic programming one, which is typically NP-Hard. In practice, a good approximation of the optimal solution could be obtained through the adoption of a closed-form expression of the performance measure in steady-state conditions. Once we have summarized the drawbacks of such optimization strategy, we address two novel optimization approaches. The first one derives from Gokbayrak and Cassandras and is based on the minimization over the discrete constraint set using an estimate of the gradient, obtained through a "relaxed continuous extension" of the performance measure. The computation of the gradient estimation is based on infinitesimal perturbation analysis (IPA). Neither closed forms of the performance measures, nor additional feedbacks concerning the state of the system and very mild assumptions about the stochastic environment are requested. The second one is the main contribution of the present work, and is based on an open-loop feedback control (OLFC) strategy, aimed at providing optimal reallocation strategies as functions of the state of the network. The optimization approach leads us to a functional optimization problem, and we investigate the adoption of a neural network-based technique, in order to approximate its solution. As is shown in the simulation results, we obtain near-optimal reallocation strategies with a small real time computational effort and avoid the suboptimal transient periods introduced by the IPA gradient descent algorithm.

There are several QoS technologies available for an Internet service provider to manage guaranteed performance services in a telecommunication network. The study of the possible interactions between the different QoS technologies is currently an open area of research often called, in the literature, QoS mapping. A QoS mapping problem arises when different encapsulation formats are employed to support a fixed QoS with different transport technologies (e.g., ATM, IP). We face this problem in terms of bandwidth management. We adopt a novel control scheme that does not need any closed-form formula for the performance metric and that is able to react to traffic changes.

The paper presents a general framework and some possible architecture to get a feasible QoS (quality of service) mapping among network portions that use different technologies to provide a fixed service level to the terminal users. In general, there is a strong need to have "communication" among portions implementing different QoS technologies: a proper architecture, functionalities and protocols should be defined. The current work, after stating the framework, tries to propose some architectural solutions and shows a preliminary performance investigation concerning mapping of a specific QoS parameter as bit loss among an ATM and an IP portion.