Publication: Veri zehirleme saldırılarının üretici modeller kullanılarak engellenmesi
Abstract
Teknolojinin hızla gelişmesi ile birlikte makine öğrenmesi yöntemleri siber güvenlik alanında da kullanılmaya başlanmış ve bir çok güvenlik önlemi kapsamında kullanılan sistemler makine öğrenmesi yöntemleri ile geliştirilmiştir. Ancak saldırıların bu yöntemler ile daha başarılı bir şekilde engellenmesi veya tespit edilmesi siber saldırganları bu yöntemleri atlatmaya yöneltmiştir. Bu kapsamda makine öğrenmesi algoritmalarının güvenliği diğer alanlara kıyasla siber güvenlik alanında daha büyük önem kazanmaktadır. Bu çalışma kapsamında makine öğrenmesi yöntemlerinden sınıflandırma modelinin eğitim aşamasından önce eğitim veri kümesinin saldırgan tarafından manipüle edildiği ve bu yolla modelin güvenliğinin ihlal edildiği bir veri zehirleme saldırısı gerçekleştirilmiştir. Ayrıca çalışma kapsamında bu saldırıya karşı makine öğrenmesinin sınıflandırma metodunu daha dayanıklı hale getirecek bir savunma mekanizması önerilmiştir. Savunma mekanizmasında üretici bir model olan oto-kodlayıcı modeli kullanılmış ve sınıflandırma metodunun eğitileceği eğitim veri kümesinin oto-kodlayıcı modeli tarafından tekrar üretilmesi ve sınıflandırma modelinin oto-kodlayıcı modeli tarafından tekrar üretilen veri ile eğitilmesi sağlanmıştır.
With the rapid development of technology, machine learning methods have also started to be used in the field of cyber security, and systems used in many security measures have been developed with machine learning methods. However, the successful prevention or detection of attacks with these methods has led cyber attackers to circumvent these methods. In this context, the security of machine learning algorithms becomes more important in the field of cyber security compared to other areas. Within the scope of this study, a data poisoning attack was carried out before the training phase of the classification model, one of the machine learning methods, and training set was manipulated by the attacker and the security of the model was violated in this way. Also, a defense mechanism is proposed to make the classification method of machine learning more robust to this attack within the scope of the study. The auto-encoder model, which is a generative model was used in the defense mechanism and the training data set, in which the classification method would be trained, was reproduced by the auto-encoder model and the classification model was trained with the data produced by the auto-coder model.
With the rapid development of technology, machine learning methods have also started to be used in the field of cyber security, and systems used in many security measures have been developed with machine learning methods. However, the successful prevention or detection of attacks with these methods has led cyber attackers to circumvent these methods. In this context, the security of machine learning algorithms becomes more important in the field of cyber security compared to other areas. Within the scope of this study, a data poisoning attack was carried out before the training phase of the classification model, one of the machine learning methods, and training set was manipulated by the attacker and the security of the model was violated in this way. Also, a defense mechanism is proposed to make the classification method of machine learning more robust to this attack within the scope of the study. The auto-encoder model, which is a generative model was used in the defense mechanism and the training data set, in which the classification method would be trained, was reproduced by the auto-encoder model and the classification model was trained with the data produced by the auto-coder model.