Classificação de conteúdo malicioso baseado em Floresta de Caminhos Ótimos
Data
2016-05-19
Autores
Fernandes, Dheny [UNESP]
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
O advento da Internet trouxe amplos benefícios nas áreas de comunicação, entretenimento, compras, relações sociais, entre outras. Entretanto, várias ameaças começaram a surgir nesse cenário, levando pesquisadores a criar ferramentas para lidar com elas. Spam, malwares, con- teúdos maliciosos, pishing, fraudes e falsas URLs são exemplos de ameaças. Em contrapartida, sistemas antivírus, firewalls e sistemas de detecção e prevenção de intrusão são exemplos de ferramentas de combate às tais ameaças. Principalmente a partir de 2010, encabeçado pelo malware Stuxnet, as ameaças tornaram-se muito mais complexas e persistentes, fazendo com que as ferramentas até então utilizadas se tornassem obsoletas. O motivo é que tais ferra- mentas, baseadas em assinaturas e anomalias, não conseguem acompanhar tanto a velocidade de desenvolvimento das ameaças quanto sua complexidade. Desde então, pesquisadores têm voltado suas atenções a métodos mais eficazes para se combater ciberameaças. Nesse contexto, algoritmos de aprendizagem de máquina estão sendo explorados na busca por soluções que analisem em tempo real ameaças provenientes da internet. Assim sendo, este trabalho tem como objetivo analisar o desempenho dos classificadores baseados em Floresta de Caminhos Ótimos, do inglês Optimum-path Forest (OPF), comparando-os com os demais classificadores do estado-da-arte. Para tanto, serão analisados dois métodos de extração de características: um baseado em tokens e o outro baseado em Ngrams, sendo N igual a 3. De maneira geral, o OPF mais se destacou no não bloqueio de mensagens legítimas e no tempo de treinamento. Em algumas bases a quantidade de spam corretamente classificada também foi alta. A versão do OPF que utiliza grafo completo foi melhor, apesar de que em alguns casos a versão com grafo knn se sobressaiu. Devido às exigências atuais em questões de segurança, o OPF, pelo seu rápido tempo de treinamento, pode ser melhorado em sua eficácia visando uma aplicação real. Em relação aos métodos de extração de características, 3gram foi superior, melhorando os resultados obtidos pelo OPF.
The advent of Internet has brought widespread benefits in the areas of communication, entertainment, shopping, social relations, among others. However, several threats began to emerge in this scenario, leading researchers to create tools to deal with them. Spam, malware, malicious content, phishing, fraud and false URLs are some examples of these threats. In contrast, anti-virus systems, firewalls and intrusion detection and prevention systems are examples of tools to combat such threats. Especially since 2010, headed by the Stuxnet malware, threats have become more complex and persistent, making the tools previously used became obsolete. The reason is that such tools based on signatures and anomalies can not follow both the speed of development of the threats and their complexity. Since then, researchers have turned their attention to more effective methods to combat cyber threats. In this context, machine learning algorithms are being exploited in the search for solutions to analyze real-time threats from the internet. Therefore, this study aims to analyze the performance of classifiers based on Optimum-path Forest, OPF, comparing them with the other state-of-the-art classifiers. To do so, two features extraction methods will be analyzed: one based on tokens and other based on Ngrams, considering N equal 3. Overall, OPF stood out in not blocking legitimate messages and training time. In some bases the amount of spam classified correctly was high as well. The version that uses complete graph was better, although in some cases the version that makes use of knn graph outperformed it. Due to the current demands on security issues, OPF, considering its fast training time, can be improved in its effectiveness aiming at a real application. In relation to feature extraction methods, 3gram was better, improving OPF’s results.
The advent of Internet has brought widespread benefits in the areas of communication, entertainment, shopping, social relations, among others. However, several threats began to emerge in this scenario, leading researchers to create tools to deal with them. Spam, malware, malicious content, phishing, fraud and false URLs are some examples of these threats. In contrast, anti-virus systems, firewalls and intrusion detection and prevention systems are examples of tools to combat such threats. Especially since 2010, headed by the Stuxnet malware, threats have become more complex and persistent, making the tools previously used became obsolete. The reason is that such tools based on signatures and anomalies can not follow both the speed of development of the threats and their complexity. Since then, researchers have turned their attention to more effective methods to combat cyber threats. In this context, machine learning algorithms are being exploited in the search for solutions to analyze real-time threats from the internet. Therefore, this study aims to analyze the performance of classifiers based on Optimum-path Forest, OPF, comparing them with the other state-of-the-art classifiers. To do so, two features extraction methods will be analyzed: one based on tokens and other based on Ngrams, considering N equal 3. Overall, OPF stood out in not blocking legitimate messages and training time. In some bases the amount of spam classified correctly was high as well. The version that uses complete graph was better, although in some cases the version that makes use of knn graph outperformed it. Due to the current demands on security issues, OPF, considering its fast training time, can be improved in its effectiveness aiming at a real application. In relation to feature extraction methods, 3gram was better, improving OPF’s results.
Descrição
Palavras-chave
Segurança em redes de computadores, Floresta de Caminhos Ótimos, Spam, Classificação, Aprendizado de máquina, Computer network security, Optimum-path Forest, Classification, Machine learning