| Summary: | Modern corporations face nowadays multiple threats within their networks. In an era where companies are tightly dependent on information, these threats can seriously compromise the safety and integrity of sensitive data. Unauthorized access and illicit programs comprise a way of penetrating the corporate networks, able to traversing and propagating to other terminals across the private network, in search of confidential data and business secrets. The efficiency of traditional security defenses are being questioned with the number of data breaches occurred nowadays, being essential the development of new active monitoring systems with artificial intelligence capable to achieve almost perfect detection in very short time frames. However, network monitoring and storage of network activity records are restricted and limited by legal laws and privacy strategies, like encryption, aiming to protect the confidentiality of private parties. This dissertation proposes methodologies to infer behavior patterns and disclose anomalies from network traffic analysis, detecting slight variations compared with the normal profile. Bounded by network OSI layers 1 to 4, raw data are modeled in features, representing network observations, and posteriorly, processed by machine learning algorithms to classify network activity. Assuming the inevitability of a network terminal to be compromised, this work comprises two scenarios: a self-spreading force that propagates over internal network and a data exfiltration charge which dispatch confidential info to the public network. Although features and modeling processes have been tested for these two cases, it is a generic operation that can be used in more complex scenarios as well as in different domains. The last chapter describes the proof of concept scenario and how data was generated, along with some evaluation metrics to perceive the model’s performance. The tests manifested promising results, ranging from 96% to 99% for the propagation case and 86% to 97% regarding data exfiltration.
|
|---|