Query by Image Examination: Classification of Digital Image-Based Forensics Using Deep Learning Methods

İlker Kara

Article Details

Query by Image Examination: Classification of Digital Image-Based Forensics Using Deep Learning Methods

Authors:

İlker Kara

Publisher:

Istanbul University

Main Topics:

Bilgisayar Yazılımı

Journal:

Acta Infologica Vol.7 Issue 2, pp.348-359

Document Type:

Article

Publication Date:

12/2023

Document Language:

English

Document ID (Org No.):

2001312784

pISSN:

2602-3563

Keywords:

Abstract:

 The continuous increase in the use of information systems and online services has also spurred the forensic examination of digital and image data, which serves as the primary platform for information transfer. In particular, according to the latest reports, the examination of the images obtained from all kinds of recording devices that have the quality of evidence as a result of the forensic case and that can provide the clarification of the incident and the detection of the criminal elements are becoming a critical problem due to the huge amount of data. Our contribution in this study is two-folded. First, we present a new approach that classifies digital images into eight different crime categories using six different models. Second, we have created a new dataset for the classification of crimes and opened it to the public. Throughout the study, we have used our new dataset which has a total of 15,065 image samples from 8 different crime categories including Bet, ChildAbuse, Credit Card and Banking, Drugs, Frightening, Knives, Pornographic and Weapons. In this study, six different models were used to classify crime images. The CNN model was developed by us and five other models used for transfer learning. Pre-trained network model parameters VGG16, VGG19, Xception Model, InceptionResNetV2 and NASNetLarge were used for crime image classification tasks. In addition, the performance of these models is compared using test accuracy and time metrics. Resultly, we achieved prediction accuracy of up to 89.74 using the NASNetLarge model.

 The continuous increase in the use of information systems and online services has also spurred the forensic examination of digital and image data, which serves as the primary platform for information transfer. In particular, according to the latest reports, the examination of the images obtained from all kinds of recording devices that have the quality of evidence as a result of the forensic case and that can provide the clarification of the incident and the detection of the criminal elements are becoming a critical problem due to the huge amount of data. Our contribution in this study is two-folded. First, we present a new approach that classifies digital images into eight different crime categories using six different models. Second, we have created a new dataset for the classification of crimes and opened it to the public. Throughout the study, we have used our new dataset which has a total of 15,065 image samples from 8 different crime categories including Bet, ChildAbuse, Credit Card and Banking, Drugs, Frightening, Knives, Pornographic and Weapons. In this study, six different models were used to classify crime images. The CNN model was developed by us and five other models used for transfer learning. Pre-trained network model parameters VGG16, VGG19, Xception Model, InceptionResNetV2 and NASNetLarge were used for crime image classification tasks. In addition, the performance of these models is compared using test accuracy and time metrics. Resultly, we achieved prediction accuracy of up to 89.74 using the NASNetLarge model.

 The continuous increase in the use of information systems and online services has also spurred the forensic examination of digital and image data, which serves as the primary platform for information transfer. In particular, according to the latest reports, the examination of the images obtained from all kinds of recording devices that have the quality of evidence as a result of the forensic case and that can provide the clarification of the incident and the detection of the criminal elements are becoming a critical problem due to the huge amount of data. Our contribution in this study is two-folded. First, we present a new approach that classifies digital images into eight different crime categories using six different models. Second, we have created a new dataset for the classification of crimes and opened it to the public. Throughout the study, we have used our new dataset which has a total of 15,065 image samples from 8 different crime categories including Bet, ChildAbuse, Credit Card and Banking, Drugs, Frightening, Knives, Pornographic and Weapons. In this study, six different models were used to classify crime images. The CNN model was developed by us and five other models used for transfer learning. Pre-trained network model parameters VGG16, VGG19, Xception Model, InceptionResNetV2 and NASNetLarge were used for crime image classification tasks. In addition, the performance of these models is compared using test accuracy and time metrics. Resultly, we achieved prediction accuracy of up to 89.74 using the NASNetLarge model.

 Bilgi sistemlerinin ve çevrimiçi hizmetlerin kullanımındaki sonsuz artış, bilgi aktarımı için temel platformlardan biri olan dĳital ve görüntü içeren verilerin adli incelemelerini de tetiklemiştir. Adli görüntü inceleme temel olarak bilimsel yöntemlerin ve adli inceleme yazılımlar kullanılarak ilgili görüntüler hakkında delil oluşturulmasını sağlayan bilimsel bir disiplindir. Özellikle, son raporlara göre, adli vaka sonucunda delil niteliği taşıyan ve olayın aydınlanmasını sağlayabilecek her türlü kayıt cihazından elde edilmiş görüntülerin incelenmesi ve suç unsuru olanlarının tespiti artan veri miktarı nedeniyle giderek büyük bir problem haline gelmektedir. Bu çalışmada katkımız iki katkı sunmaktadır. İlk olarak dĳital görüntülerin altı farklı model kullanarak sekiz farklı suç kategorisi olarak sınıflandıran yeni bir yaklaşım sunuyor. İkincisi, suçların sınıflandırılması için yeni bir veri kümesinin oluşturarak paylaşıma sunuyor. Çalışma boyunca, Bet, ChildAbuse, kredi kartı ve bankacılık, uyuşturucu, korkutucu, bıçak, pornografik ve silah dâhil olmak üzere 8 farklı suç Kategorisine ait toplam 15.065 görüntü örneğini kapsayan yeni veri setimizi kullanıldı. Suç görüntülerini sınıflandırmak için bu çalışmada 6 farklı model kullanılmıştır. CNN modeli kendimiz ve öğrenmeyi ince ayarlara aktarmak için kullanılan diğer beş model tarafından yaratılmıştır. Görüntü sınıflandırma görevleri için VGG16, VGG19, Xception modeli, InceptionResNetV2 ve NASNetLarge önceden eğitilmiş ağ modeli parametreleri kullanıldı. Ayrıca, bu modellerin performansı test doğruluğu ve zaman ölçümleri kullanılarak karşılaştırılır. Sonuçlar, NASNetLarge modeli kullanarak 89.74e kadar tahmin doğruluğu elde edilmiştir.

Article Details

Query by Image Examination: Classification of Digital Image-Based Forensics Using Deep Learning Methods

Authors:

İlker Kara

Publisher:

Istanbul University

Main Topics:

Bilgisayar Yazılımı

Journal:

Acta Infologica Vol.7 Issue 2, pp.348-359

Document Type:

Article

Publication Date:

12/2023

Document Language:

English

Document ID (Org No.):

2001312784

pISSN:

2602-3563

image processing

deep learning method

image classification

data mining

forensic investigationGörüntü işleme

derin öğrenme yöntemi

görüntü sınıflandırması

veri madenciliği

adli inceleme

Issues

Article Details

Authors: İlker Kara

Publisher: Istanbul University

Main Topics: Bilgisayar Yazılımı

Journal: Acta Infologica Vol.7 Issue 2, pp.348-359

Document Type: Article

Publication Date: 12/2023

Document Language: English

Document ID (Org No.): 2001312784

pISSN: 2602-3563

Issues

Authors:

İlker Kara

Publisher:

Istanbul University

Main Topics:

Bilgisayar Yazılımı

Journal:

Acta Infologica Vol.7 Issue 2, pp.348-359

Document Type:

Article

Publication Date:

12/2023

Document Language:

English

Document ID (Org No.):

2001312784

pISSN:

2602-3563