Globally, the number of internet users increases every year. As a matter of fact, we use technological devices to surf the internet, for online shopping, or just to relax and keep our relationships by spending time on social networks. By doing any of those actions, we release information that can be used in many ways, such as targeted advertising via cookies but also abused by malicious users for scams or theft. On the other hand, many detection systems have been developed with the aim to counteract malicious actions. In particular, special attention has been paid to the malware, designed to perpetrate malicious actions inside software systems and widespread through internet networks or e-mail messages. In this paper, we propose a deep learning model aimed to detect ransomware. We propose a set of experiments aimed to demonstrate that the proposed method obtains good accuracy during the training and test phases across a dataset of over 15,000 elements. Moreover, to improve our results and interpret the output obtained from the models, we have also exploited the Gradient-weighted Class Activation Mapping.
Explainable Ransomware Detection with Deep Learning Techniques
Mercaldo F.;Santone A.
2023-01-01
Abstract
Globally, the number of internet users increases every year. As a matter of fact, we use technological devices to surf the internet, for online shopping, or just to relax and keep our relationships by spending time on social networks. By doing any of those actions, we release information that can be used in many ways, such as targeted advertising via cookies but also abused by malicious users for scams or theft. On the other hand, many detection systems have been developed with the aim to counteract malicious actions. In particular, special attention has been paid to the malware, designed to perpetrate malicious actions inside software systems and widespread through internet networks or e-mail messages. In this paper, we propose a deep learning model aimed to detect ransomware. We propose a set of experiments aimed to demonstrate that the proposed method obtains good accuracy during the training and test phases across a dataset of over 15,000 elements. Moreover, to improve our results and interpret the output obtained from the models, we have also exploited the Gradient-weighted Class Activation Mapping.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.