eXplainable and Reliable Against Adversarial Machine Learning in Data Analytics

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