User profiles’ image clustering for digital investigations

Sharing images on Social Network (SN) platforms is one of the most widespread behaviors which may cause privacy-intrusive and illegal content to be widely distributed. Clustering the images shared through SN platforms according to the acquisition cameras embedded in smartphones is regarded as a significant task in forensic investigations of cybercrimes. The Sensor Pattern Noise (SPN) caused by camera sensor imperfections due to the manufacturing process has been proved to be an effective and robust camera fingerprint that can be used for several tasks, such as digital evidence analysis, smartphone fingerprinting and user profile linking as well. Clustering the images uploaded by users on their profiles is a way of fingerprinting the camera sources and it is considered a challenging task since users may upload different types of images, i.e., the images taken by users’ smartphones (taken images) and single images from different sources, cropped images, or generic images from the Web (shared images). The shared images make a perturbation in the clustering task, as they do not usually present sufficient characteristics of SPN of their related sources. Moreover, they are not directly referable to the user’s device so they have to be detected and removed from the clustering process. In this paper, we propose a user profiles’ image clustering method without prior knowledge about the type and number of the camera sources. The hierarchical graph-based method clusters both types of images, taken images and shared images. The strengths of our method include overcoming large-scale image datasets, the presence of shared images that perturb the clustering process and the loss of image details caused by the process of content compression on SN platforms. The method is evaluated on the VISION dataset, which is a public benchmark including images from 35 smartphones. The dataset is perturbed by 3000 images, simulating the shared images from different sources except for users’ smartphones. Experimental results confirm the robustness of the proposed method against perturbed datasets and its effectiveness in the image clustering.