Predicting Stealthy Watermarks in Files Using Deep Learning

Digital evidence is a fundamental element in cyber-forensics and judicial processes. However, the work of forensic examiners is becoming more challenging as the volume digital content and files increases. In this paper, we use machine learning especially deep learning to detect stealthy watermarks in various types of files. We use a black box approach which is different from current steganographic and cryptographic methods to find patterns of candidate file locations for hidden data We studied Deep Neural Networks (DNN) to predict stealthy watermarks in files using the deep learning implementation (DL4J) and Multilayer Perceptron (MLP) algorithms as implemented in Weka. We evaluated MLP models by altering the number of neurons and hidden layers while the DL4J models were evaluated by varying the number of dense layers and nodes. For the MLP models, DOCX & PPTX singleton models predicted stealthy watermarks in files with predictive accuracies ranging from 47.5% to 100%; JPEG singleton models registered predictive accuracies ranging from 35% to 65%. Comparatively, HYBRID3 models had predictive accuracies ranging from 42.5% to 95% while HYBRID_OOXML had predictive accuracies of 47.5% to 100%. However, JPEG_DOCX had predictive accuracies 47.5% to 97.5% while JPEG_PPTX had predictive accuracies of 40% to 85%. Furthermore for DL4J models, we only generated HYBRID3 models, which predicted stealthy watermarks in DOCX files with predictive accuracies 100%. The HYBRID3 DL4J model predicted stealthy watermarks in PPTX with predictive accuracies ranging from 55% to 82 % while in JPEG, the predictive accuracies from 50% to 52.5%. The major finding with deep learning also revealed improvements in prediction of stealthy watermarks in PPTX files using DL4J models.