Leveraging Deep Learning for Identification of Illicit Images in Digital Forensic Investigations

Abstract

Background: The distribution of illicit material such as obscene images and child sexual abuse (CSA) content constitutes a serious crime in numerous jurisdictions worldwide. The identification and detection of such content is a crucial component of forensic investigations, and is integral to the apprehension and prosecution of offenders. Traditional methods for identifying such material are predominantly manual, requiring expert review, and are consequently time-consuming and susceptible to human error. Deep learning, with its proficiency in discerning complex patterns and features in large-scale data, offers a promising alternative for the automated and accurate detection of obscene and CSA content. This research advances a deep learning model for detecting obscene images and CSA in digital forensic evidence. Methods: A dataset of 3000 obscene and 3000 non-obscene images was compiled, with the obscene images sourced from a social sharing platform, Reddit. Images were classified as obscene if they portrayed sexual organs or activities for the primary purpose of eliciting sexual arousal. A convolutional neural network (CNN) based deep learning model was then developed to detect the obscene content. The efficacy of the proposed model was compared with existing methodologies using the NPDI benchmark dataset. Pre-trained CNNs were used to extract feature vectors from the images which were subsequently used in conjunction with a neural network classifier to categorise the images. To detect CSA, the UTKFace dataset was employed to identify minors in images, using a lightweight CNN model with skip connections to recognise juvenile faces. Results: The proposed model for obscene content detection demonstrated a robust performance, achieving an accuracy of 99.8% and 99.4% respectively in the training and testing segments of the NPDI dataset. This model has potential applicability to both image and video files. Meanwhile, the model for identifying minors achieved an accuracy of 99.6% and 99.0% in the training and testing segments of the UTKFace dataset respectively. Conclusion: The findings of this study underscore the efficacy and high performance of the deep learning models proposed for the detection of obscene images and CSA content. These results highlight the potential for these models to be employed in digital forensic investigations for automated content detection, which would significantly advance efforts in combatting the distribution of illicit content.