• Privacy verification of photoDNA based on machine learning

      Nadeem, Muhammad Shahroz; Franqueira, Virginia N. L.; Zhai, Xiaojun; University of Derby, College of Engineering and Technology; University of Essex, School of Computer Science and Electronic Engineering (The Institution of Engineering and Technology (IET), 2019-10-09)
      PhotoDNA is a perceptual fuzzy hash technology designed and developed by Microsoft. It is deployed by all major big data service providers to detect Indecent Images of Children (IIOC). Protecting the privacy of individuals is of paramount importance in such images. Microsoft claims that a PhotoDNA hash cannot be reverse engineered into the original image; therefore, it is not possible to identify individuals or objects depicted in the image. In this chapter, we evaluate the privacy protection capability of PhotoDNA by testing it against machine learning. Specifically, our aim is to detect the presence of any structural information that might be utilized to compromise the privacy of the individuals via classification. Due to the widespread usage of PhotoDNA as a deterrent to IIOC by big data companies, ensuring its ability to protect privacy would be crucial. In our experimentation, we achieved a classification accuracy of 57.20%.This result indicates that PhotoDNA is resistant to machine-learning-based classification attacks.
    • A survey of deep learning solutions for multimedia visual content analysis.

      Nadeem, Muhammad Shahroz; Franqueira, Virginia N. L.; Zhai, Xiaojun; Kurugollu, Fatih; University of Derby; University of Essex (IEEE, 2019-06-24)
      The increasing use of social media networks on handheld devices, especially smartphones with powerful built-in cameras, and the widespread availability of fast and high bandwidth broadband connections, added to the popularity of cloud storage, is enabling the generation and distribution of massive volumes of digital media, including images and videos. Such media is full of visual information and holds immense value in today’s world. The volume of data involved calls for automated visual content analysis systems able to meet the demands of practice in terms of efficiency and effectiveness. Deep Learning (DL) has recently emerged as a prominent technique for visual content analysis. It is data-driven in nature and provides automatic end-to-end learning solutions without the need to rely explicitly on predefined handcrafted feature extractors. Another appealing characteristic of DL solutions is the performance they can achieve, once the network is trained, under practical constraints. This paper identifies eight problem domains which require analysis of visual artefacts in multimedia. It surveys the recent, authoritative, and best performing DL solutions and lists the datasets used in the development of these deep methods for the identified types of visual analysis problems. The paper also discusses the challenges that DL solutions face which can compromise their reliability, robustness, and accuracy for visual content analysis.