Malware Detection in Forensic Memory Dumps: The Use of Deep Meta-Learning Models

Yalçın Özkan

Makale detayları

Malware Detection in Forensic Memory Dumps: The Use of Deep Meta-Learning Models

Yazarlar:

Yalçın Özkan

Yayıncı:

Istanbul University

Ana Başlıklar:

Yazılım Mühendisliği (Diğer)

Dergi:

Acta Infologica Vol.7 Issue 1, pp.165-172

Belge Tipi:

Makale

Yayın Tarihi:

2/2024

Belge Dili:

İngilizce

Document ID (Org No.):

2001309400

pISSN:

2602-3563

Anahtar kelimeler:

Özet:

 The present study aimed to design a high-performance deep meta-learning model that could be utilized in classification predictions using forensic memory datasets and propose a framework that would ensure the generalization and consistency of the predictions with the help of this model. To achieve this aim, a dataset containing malware and obtained from forensic memory dumps was addressed. First, it was subjected to the classification process with a deep learning algorithm, and a predictive model was acquired. The predictive model was found to have an accuracy metric of 98.25. In addition to this finding, a meta-learning model consisting of five different models with the same hyperparameters was created. The accuracy of the obtained meta-model was computed as 97.69. With the thought that this model would reduce the prediction variance and thus the predictive model could be generalized, it was ensured to be run 5 times in a row. As a result of this process, the prediction variance, indicating a very small change, was calculated as 0.000012. Accordingly, considering the acquired performance value, it can be determined that high performance is achieved in malware detection, and thus what hyperparameters ensure success can be revealed. If deep learning methods are used as a single model, the problem is that the variance between the predictions is large due to its stochastic structure. To avoid such drawbacks, a deep meta-learning model using the same parameters was designed instead of a deep learning model comprising a single model, and considerably smaller variance values were achieved, thus providing generalized and consistent predictions.

 The present study aimed to design a high-performance deep meta-learning model that could be utilized in classification predictions using forensic memory datasets and propose a framework that would ensure the generalization and consistency of the predictions with the help of this model. To achieve this aim, a dataset containing malware and obtained from forensic memory dumps was addressed. First, it was subjected to the classification process with a deep learning algorithm, and a predictive model was acquired. The predictive model was found to have an accuracy metric of 98.25. In addition to this finding, a meta-learning model consisting of five different models with the same hyperparameters was created. The accuracy of the obtained meta-model was computed as 97.69. With the thought that this model would reduce the prediction variance and thus the predictive model could be generalized, it was ensured to be run 5 times in a row. As a result of this process, the prediction variance, indicating a very small change, was calculated as 0.000012. Accordingly, considering the acquired performance value, it can be determined that high performance is achieved in malware detection, and thus what hyperparameters ensure success can be revealed. If deep learning methods are used as a single model, the problem is that the variance between the predictions is large due to its stochastic structure. To avoid such drawbacks, a deep meta-learning model using the same parameters was designed instead of a deep learning model comprising a single model, and considerably smaller variance values were achieved, thus providing generalized and consistent predictions.

 The present study aimed to design a high-performance deep meta-learning model that could be utilized in classification predictions using forensic memory datasets and propose a framework that would ensure the generalization and consistency of the predictions with the help of this model. To achieve this aim, a dataset containing malware and obtained from forensic memory dumps was addressed. First, it was subjected to the classification process with a deep learning algorithm, and a predictive model was acquired. The predictive model was found to have an accuracy metric of 98.25. In addition to this finding, a meta-learning model consisting of five different models with the same hyperparameters was created. The accuracy of the obtained meta-model was computed as 97.69. With the thought that this model would reduce the prediction variance and thus the predictive model could be generalized, it was ensured to be run 5 times in a row. As a result of this process, the prediction variance, indicating a very small change, was calculated as 0.000012. Accordingly, considering the acquired performance value, it can be determined that high performance is achieved in malware detection, and thus what hyperparameters ensure success can be revealed. If deep learning methods are used as a single model, the problem is that the variance between the predictions is large due to its stochastic structure. To avoid such drawbacks, a deep meta-learning model using the same parameters was designed instead of a deep learning model comprising a single model, and considerably smaller variance values were achieved, thus providing generalized and consistent predictions.

 Bu çalışmada adli bellek veri kümelerinden yararlanılarak, sınıflandırma öngörülerinde kullanılabilecek yüksek performanslı bir derin meta öğrenme modelinin tasarlanması ve bu model yardımıyla öngörülerin genelleştirme ve tutarlılığını sağlayacak bir çerçevenin önerilmesi amaçlanmaktadır. Bu amaca ulaşabilmek için, kötü amaçlı yazılımları içeren ve adli bellek dökümlerinden elde edilen bir veri kümesi ele alınarak önce derin öğrenme algoritması ile sınıflandırma sürecine tabi tutuldu ve bir öngörü modeli elde edildi. Öngörü modelinin 98,25lik bir doğruluk metriğine sahip olduğu görülmüştür. Bu bulgunun yanı sıra, aynı hiper parametrelere sahip 5 ayrı modelden oluşan bir meta öğrenme modeli oluşturulmuştur. Elde edilen meta modelin doğruluğu 97,69 olarak hesaplandı. Bu modelin öngörü varyansını azaltacağı ve böylece öngörü modelini genelleştirilebileceği düşüncesiyle ardı ardına 5 kez çalıştırılması sağlandı. Bu işlem sonucunda çok küçük bir değişime işaret eden öngörü varyansı 0,000012 olarak hesaplandı. Sonuç olarak, elde edilen performans değeri göz önüne alındığında, kötü amaçlı yazılım tespitinde yüksek bir performansın elde edildiği ve böylece başarıyı sağlayan hiper parametrelerin neler olduğu belirlenebilmektedir. Derin öğrenme yöntemlerinin tek model olarak kullanılması durumunda, stokastik bir yapıya sahip olması nedeniyle öngörüler arasındaki varyansın büyük olması sorunuyla karşılaşılmaktadır. Bu tür sakıncaları önlemek üzere, tek modelden oluşan derin öğrenme modeli yerine, aynı parametreleri kullanan bir derin meta öğrenme modeli tasarlanarak çok daha küçük varyans değerlerine ulaşılmış, böylece genelleştirilmiş ve tutarlı öngörüler üretilmesi sağlanmıştır.

Makale detayları

Malware Detection in Forensic Memory Dumps: The Use of Deep Meta-Learning Models

Yazarlar:

Yalçın Özkan

Yayıncı:

Istanbul University

Ana Başlıklar:

Yazılım Mühendisliği (Diğer)

Dergi:

Acta Infologica Vol.7 Issue 1, pp.165-172

Belge Tipi:

Makale

Yayın Tarihi:

2/2024

Belge Dili:

İngilizce

Document ID (Org No.):

2001309400

pISSN:

2602-3563

Forensic memory

Cyber security

Deep learning

Meta-learningAdli bellek

Siber güvenlik

Derin öğrenme

Meta öğrenme

Sayı

Makale detayları

Yazarlar: Yalçın Özkan

Yayıncı: Istanbul University

Ana Başlıklar: Yazılım Mühendisliği (Diğer)

Dergi: Acta Infologica Vol.7 Issue 1, pp.165-172

Belge Tipi: Makale

Yayın Tarihi: 2/2024

Belge Dili: İngilizce

Document ID (Org No.): 2001309400

pISSN: 2602-3563

Sayı

Yazarlar:

Yalçın Özkan

Yayıncı:

Istanbul University

Ana Başlıklar:

Yazılım Mühendisliği (Diğer)

Dergi:

Acta Infologica Vol.7 Issue 1, pp.165-172

Belge Tipi:

Makale

Yayın Tarihi:

2/2024

Belge Dili:

İngilizce

Document ID (Org No.):

2001309400

pISSN:

2602-3563