The Internet of Things (IoT) became established during the last decade as an emerging technology with considerable potentialities and applicability. Its paradigm of everything connected together penetrated the real world, with smart devices located in several daily appliances. Such intelligent objects are able to communicate autonomously through already existing network infrastructures, thus generating a more concrete integration between real world and computer-based systems. On the downside, the great benefit carried by the IoT paradigm in our life brings simultaneously severe security issues, since the information exchanged among the objects frequently remains unprotected from malicious attackers. The paper at hand proposes COSMOS (Collaborative, Seamless and Adaptive Sentinel for the Internet of Things), a novel sentinel to protect smart environments from cyber threats. Our sentinel shields the IoT devices using multiple defensive rings, resulting in a more accurate and robust protection. Additionally, we discuss the current deployment of the sentinel on a commodity device (i.e., Raspberry Pi). Exhaustive experiments are conducted on the sentinel, demonstrating that it performs meticulously even in heavily stressing conditions. Each defensive layer is tested, reaching a remarkable performance, thus proving the applicability of COSMOS in a distributed and dynamic scenario such as IoT. With the aim of easing the enjoyment of the proposed sentinel, we further developed a friendly and ease-to-use COSMOS App, so that end-users can manage sentinel(s) directly using their own devices (e.g., smartphone).

Sandboxing has been used regularly to analyze software samples and determine if these contain suspicious properties or behaviors. Even if sandboxing is a powerful technique to perform malware analysis, it requires that a malware analyst performs a rigorous analysis of the results to determine the nature of the sample: goodware or malware. This paper proposes two machine learning models able to classify samples based on signatures and permissions obtained through Cuckoo sandbox, Androguard and VirusTotal. The developed models are also tested obtaining an acceptable percentage of correctly classified samples, being in this way useful tools for a malware analyst. A proposal of architecture for an IoT sentinel that uses one of the developed machine learning model is also showed. Finally, different approaches, perspectives, and challenges about the use of sandboxing and machine learning by security teams in State security agencies are also shared. El sandboxing ha sido usado de manera regular para analizar muestras de software y determinar si estas contienen propiedades o comportamientos sospechosos. A pesar de que el sandboxing es una técnica poderosa para desarrollar análisis de malware, esta requiere que un analista de malware desarrolle un análisis riguroso de los resultados para determinar la naturaleza de la muestra: goodware o malware. Este artículo propone dos modelos de aprendizaje automáticos capaces de clasificar muestras con base a un análisis de firmas o permisos extraídos por medio de Cuckoo sandbox, Androguard y VirusTotal. En este artículo también se presenta una propuesta de arquitectura de centinela IoT que protege dispositivos IoT, usando uno de los modelos de aprendizaje automáticos desarrollados anteriormente. Finalmente, diferentes enfoques y perspectivas acerca del uso de sandboxing y aprendizaje automático por parte de agencias de seguridad del Estado también son aportados.