Machine learning (ML) algorithms are nowadays widely adopted in different contexts to perform autonomous decisions and predictions. Due to the high volume of data shared in the recent years, ML algorithms are more accurate and reliable since training and testing phases are more precise. An important concept to analyze when defining ML algorithms concerns adversarial machine learning attacks. These attacks aim to create manipulated datasets to mislead ML algorithm decisions. In this work, we propose new approaches able to detect and mitigate malicious adversarial machine learning attacks against a ML system. In particular, we investigate the Carlini-Wagner (CW), the fast gradient sign method (FGSM) and the Jacobian based saliency map (JSMA) attacks. The aim of this work is to exploit detection algorithms as countermeasures to these attacks. Initially, we performed some tests by using canonical ML algorithms with a hyperparameters optimization to improve metrics. Then, we adopt original reliable AI algorithms, either based on eXplainable AI (Logic Learning Machine) or Support Vector Data Description (SVDD). The obtained results show how the classical algorithms may fail to identify an adversarial attack, while the reliable AI methodologies are more prone to correctly detect a possible adversarial machine learning attack. The evaluation of the proposed methodology was carried out in terms of good balance between FPR and FNR on real world application datasets: Domain Name System (DNS) tunneling, Vehicle Platooning and Remaining Useful Life (RUL). In addition, a statistical analysis was performed to improve the robustness of the trained models, including evaluating their performance in terms of runtime and memory consumption.

In the era of Industry 4.0, the use of Artificial Intelligence (AI) is widespread in occupational settings. Since dealing with human safety, explainability and trustworthiness of AI are even more important than achieving high accuracy. eXplainable AI (XAI) is investigated in this paper to detect physical fatigue during manual material handling task simulation. Besides comparing global rule-based XAI models (LLM and DT) to black-box models (NN, SVM, XGBoost) in terms of performance, we also compare global models with local ones (LIME over XGBoost). Surprisingly, global and local approaches achieve similar conclusions, in terms of feature importance. Moreover, an expansion from local rules to global rules is designed for Anchors, by posing an appropriate optimization method (Anchors coverage is enlarged from an original low value, 11%, up to 43%). As far as trustworthiness is concerned, rule sensitivity analysis drives the identification of optimized regions in the feature space, where physical fatigue is predicted with zero statistical error. The discovery of such “non-fatigue regions” helps certifying the organizational and clinical decision making.

Nowadays Artificial Intelligence (AI) is bursting in many fields, including critical ones, giving rise to reliable AI, that means ensuring safety of autonomous decisions. As the false negatives may have a safety impact (e.g., in a mobility scenario, prediction of no collision, but collision in reality), the aim is to push false negatives as close to zero as possible, thus designing ‘`safety regions’' in the feature space with statistical zero error. We show here how sensitivity analysis of an eXplainable AI model drives such statistical assurance. We test and compare the proposed algorithms on two different datasets (physical fatigue and vehicle platooning) and achieve quite different conclusions in terms of results that strongly depend on the level of noise in the dataset rather than on the algorithms at hand.

Internet of Things is a widely adopted and pervasive technology, but also one of the most relevant in cyber-security, given the volume and sensitivity of shared data and the availability of affordable but insecure products. In this paper, we propose a novel cyber-threat exploiting the Message Queue Telemetry Transport (MQTT) protocol to implement a tunneling attack. In IoT networks, sensitive and critical information are exchanged between devices or external systems to perform data analysis. For this reason, a tunneling threat could be adopted by a malicious user to steal information. In this context, a tunneling system based on MQTT can be considered since this communication protocol could be allowed to pass through enterprise firewalls because it is widely adopted in this IoT world. An attacker can exploit the MQTT protocol for various purposes such as steal information or access to not-allowed websites/servers. In particular in this work, we contribute in two main points: initially we demonstrate how the proposed threat is able to encapsulate messages through the MQTT protocol, by also comparing it with other tunneling systems exploiting different communication protocols. Obtained results show that exploiting MQTT for tunneling purposes is a good choice, compared to other communication protocols, especially for payloads up to 3000 bytes. Then, we propose and validate an initial machine learning based approach able to detect the proposed MQTT tunnel, by comparing different detection algorithms tested with and without a hyperparameter optimization, in terms of accuracy, F1 score and Receiver Operating Characteristic (ROC) curve. In this case, obtained results show that some algorithms are able to identify the attack, with an accuracy exceeding 95%, while others lack of such capability.

The explosion of artificial intelligence methods has paved the way for more sophisticated smart mobility solutions. In this work, we present a multi-camera video content analysis (VCA) system that exploits a single-shot multibox detector (SSD) network to detect vehicles, riders, and pedestrians and triggers alerts to drivers of public transportation vehicles approaching the surveilled area. The evaluation of the VCA system will address both detection and alert generation performance by combining visual and quantitative approaches. Starting from a SSD model trained for a single camera, we added a second one, under a different field of view (FOV) to improve the accuracy and reliability of the system. Due to real-time constraints, the complexity of the VCA system must be limited, thus calling for a simple multi-view fusion method. According to the experimental test-bed, the use of two cameras achieves a better balance between precision (68%) and recall (84%) with respect to the use of a single camera (i.e., 62% precision and 86% recall). In addition, a system evaluation in temporal terms is provided, showing that missed alerts (false negatives) and wrong alerts (false positives) are typically transitory events. Therefore, adding spatial and temporal redundancy increases the overall reliability of the VCA system.

Out-of-distribution detection is one of the most critical issue in the deployment of machine learning. The data analyst must assure that data in operation should be compliant with the training phase as well as understand if the environment has changed in a way that autonomous decisions would not be safe anymore. The method of the paper is based on eXplainable Artificial Intelligence (XAI); it takes into account different metrics to identify any resemblance between in-distribution and out of, as seen by the XAI model. The approach is non-parametric and distributional assumption free. The validation over complex scenarios (predictive maintenance, vehicle platooning, covert channels in cybersecurity) corroborates both precision in detection and evaluation of training-operation conditions proximity.