Classifying Revenge Porn: Developing Deep Neural Networks with Facial Recognition Techniques for the Classification of Non-consensual Content on Online Hosting Platforms

Written by Paul Tuohy

This is my first official submission to my research capstone. It is a project brief which captures my current area of interest, what the problem domain is, and what I hope to achieve. I believe this to be a relative reflection of my current understanding but expect my aims and interests to deviate along the coming weeks and therefore the project title will reflect that change.

Project Title:

Classifying Revenge Porn: Developing Deep Neural Networks with Facial Recognition Techniques for the Classification of Non-consensual Content on Online Hosting Platforms.

Project brief

• Abstract:
  The ease to access and proliferate pornographic content online has placed a greater focus on an individual’s privacy and consent around content being shared online. In particular the proliferation of non-consensual ‘revenge porn’ content has become an area of concern. Manually monitoring content is impossible due to the vast amount of content available and the rate at which content can be reuploaded across multiple platforms. Therefore, developing advanced AI approaches to detect non-consensual porn are required to assist in preventing further proliferation. Unfortunately, current models are error-prone, suffer from skin-tone bias, or don’t utilise facial recognition techniques in classifying subjects in revenge porn against other pornographic content. Deep Neural Networks (DNNs) has become the best technique in computer vision processing and has great success in detecting pornography. This project aims to develop a model that combines DNNs with facial recognition techniques to produce effective classification of revenge porn.

• Background/Description:
  Non-consensual content is the unsolicited proliferation of images/videos of a subject(s) with the intended (or unintended) consequence of social harm to the subject in the content. Non-consensual content predominantly refers to revenge porn, but more broadly can include non-pornographic content such as memes displaying an individual.
  Current techniques for classifying pornographic content can be categorised into three different methods. 1) skin-based methods, which are limited by the requirement of nude-hue pixels which are not strictly limited to pornographic content; 2) feature extraction and pattern recognition methods, which include visual bag-of-words, scale invariant feature transformation (SIFT), and self-similarity; and 3) deep learning-based models, which may utilise a combination of these techniques simultaneously to classify content.

• Aim/Objectives:
  The aim of the project is to discuss the effectiveness of neural network applications for classifying revenge porn across different content formats (video and images). Further, the project will discuss the ethics of consent in relation to ownership of pornographic and non-pornographic content.

• Problem Domain:
  The process of manually classifying porn videos, in particular revenge porn videos, is a tedious task due to the vast amount pornographic and non-pornographic content, the rate at which content is uploaded, across multiple platforms, and the capabilities for users to quickly reupload content that has been taken down. The development of advanced AI techniques for classifying pornographic content has focussed predominantly on Child Sexual Abuse content and image classification (less so video formats). As an individual’s reliance on the internet grows and the need for stronger privacy around content ownership becomes more important, there is a great need for effective techniques for classifying non-consensual content. Non-consensual content can have broad negative ramifications for the subject and therefore the takedown of content should be done quickly to prevent any further harm.

• Research Questions ?

  • How effective are current porn classification techniques for different formats ?
  • Does the application of advanced AI techniques assist in the identification of revenge porn ?
  • What are the current flaws and error observed in existing models for identifying pornographic content ?

• Deliverables/Outcomes:
  The project will explore and propose a strategy for how neural networks will provide applications for assisting in classifying revenge porn across online hosting platforms.

Resources consulted:
As of 28/11/2021, I have consulted two journal articles and one conference paper:

• Gangwar, A., González-Castro, V., Alegre, E., & Fidalgo, E. (2021). AttM-CNN: Attention and metric learning based CNN for pornography, age and Child Sexual Abuse (CSA) Detection in images. Neurocomputing, 445, 81–104. https://doi.org/10.1016/j.neucom.2021.02.056

• Mohanty, M., Zhang, M., & Russello, G. (2019). A Photo Forensics-Based Prototype to Combat Revenge Porn. Proceedings - 2nd International Conference on Multimedia Information Processing and Retrieval, MIPR 2019, 5–8. https://doi.org/10.1109/MIPR.2019.00009

• Xu, W., Parvin, H., & Izadparast, H. (2020). Deep Learning Neural Network for Unconventional Images Classification. Neural Processing Letters, 52(1), 169–185. https://doi.org/10.1007/s11063-020-10238-3

Alignment of the project with specialisation:
This project aligns with my specialisation of Data Science, by exploring the application of advanced AI approaches such as deep learning neural networks, feature classification, and facial recognition, to produce solutions to real-life problems.

← Home