In this article, we study the properties of quantitative steganography detectors (estimators of the payload size) for content-adaptive steganography. In contrast to non-adaptive embedding, the estimator's bias as well as variance strongly depend on the true payload size. Initially, and depending on the image content, the estimator may not react to embedding. With increased payload size, it starts responding as the embedding changes begin to ``spill'' into regions where their detection is more reliable. We quantify this behavior with the concepts of reactive and estimable payloads. To better understand how the payload estimate and its bias depend on image content, we study a maximum likelihood estimator derived for the MiPOD model of the cover image. This model correctly predicts trends observed in outputs of a state-of-the-art deep learning payload regressor. Moreover, we use the model to demonstrate that the cover bias can be caused by a small number of ``outlier'' pixels in the cover image. This is also confirmed for the deep learning regressor on a dataset of artificial images via attribution maps.
Many areas of forensics are moving away from the notion of classifying evidence simply as a match or non-match. Instead, some use score-based likelihood ratios (SLR) to quantify the similarity between two pieces of evidence, such as a fingerprint obtained from a crime scene and a fingerprint obtained from a suspect. We apply trace-anchored score-based likelihood ratios to the camera device identification problem. We use photo-response non-uniformity (PRNU) as a camera fingerprint and one minus the normalized correlation as a similarity score. We calculate trace-anchored SLRs for 10,000 images from seven camera devices from the BOSSbase image dataset. We include a comparison between our results the universal detector method.
A new rule for modulating costs in side-informed steganography is proposed. The modulation factors of costs are determined by the minimum perturbation of the precover to quantize to the desired stego value. This new rule is contrasted with the established way of weighting costs by the difference between the rounding errors to the cover and stego values. Experiments are used to demonstrate that the new rule improves security in ternary side-informed UNIWARD in JPEG domain. The new rule arises naturally as the correct cost modulation for JPEG side-informed steganography with the “trunc” quantizer used in many portable digital imaging devices.
The purpose of this study is to prepare a source of realistic looking images in which optimal steganalysis is possible by enforcing a known statistical model on image pixels to assess the efficiency of detectors implemented using machine learning. Our goal is to answer the questions that researchers keep asking: “Are our empirical detectors close to what can be possibly detected? How much room is there for improvement?” or simply “Are we there yet?” Our goal is achieved by applying denoising to natural images to remove complex statistical dependencies introduced by processing and, subsequently, adding noise of simpler and known statistical properties that allows deriving the likelihood ratio test in a closed form. This theoretical upper bound informs us about the amount of further possible improvement. Three content-adaptive stego algorithms in the spatial domain and non-adaptive LSB matching are used to contrast the upper bound with the performance of two modern detection paradigms: a convolutional neural network and a classifier with the maxSRMd2 rich model. The short answer to the posed question is “We are much closer now but there is still non-negligible room for improvement.”