Guest Editorial

Special issue on Deep Learning in Image and Video Forensics

In recent times, deep learning models have achieved state-of-the-art performances in image classification. However, the classification and generalization ability of most achievements have highly relied on the availability of large-scale accurate labeled training data, which are time-consuming, laborious, and expensive to collect. Moreover, in the generalizing process of deep learning models, the noisy labels are the challenges for multi-category classification. Hence, it is essential to explore effective methods that can efficiently and correctly train deep models under label noise to conduct multi-category classification. This paper proposes using robust binary loss functions to train deep models under label noise to address this problem. Specifically, we suggest handling the K-class classification task by using K binary classifiers, which can be completed by a joint adoption of multi-category large margin classification approaches, e.g., Pairwise-Comparison (PC) or One-versus-All (OVA). We also theoretically demonstrate that our method is inherently tolerant to label noise by using symmetric binary loss functions in multi-category classification tasks. Moreover, we designed a truncated CCE loss to combine with the proposed losses to improve the learning ability. Finally, we test our method on three different benchmark datasets with different types of label noise. The experimental results have clearly confirmed the effectiveness of our method, which can reduce the negative effect of noisy labels and improve the generalization ability.

Many recent studies have shown that deep neural models are vulnerable to adversarial samples: images with imperceptible perturbations, for example, can fool image classifiers. In this paper, we present the first type-specific approach to generating adversarial examples for object detection, which entails detecting bounding boxes around multiple objects present in the image and classifying them at the same time, making it a harder task than against image classification. We specifically aim to attack the widely used Faster R-CNN by changing the predicted label for a particular object in an image: where prior work has targeted one specific object (a stop sign), we generalize to arbitrary objects, with the key challenge being the need to change the labels of all bounding boxes for all instances of that object type. To do so, we propose a novel method, named Pick-Object-Attack. Pick-Object-Attack successfully adds perturbations only to bounding boxes for the targeted object, preserving the labels of other detected objects in the image. In terms of perceptibility, the perturbations induced by the method are very small. Furthermore, for the first time, we examine the effect of adversarial attacks on object detection in terms of a downstream task, image captioning; we show that where a method that can modify all object types leads to very obvious changes in captions, the changes from our constrained attack are much less apparent.

Steganography is the process of hiding messages inside an object known as a carrier. The idea is establishing a covert communication channel where messages go unnoticed by observers having access to that channel. Steganalysis is dedicated to the detection of such hidden messages; these messages can be embedded in several different types of digital media, such as images, video, audio, plain text files, and covert channels. Traditional steganalysis schemes are divided into two stages; the first one involves manual extraction of high-end features, and the second one is classification using machine learning (ML) methods. Nowadays, deep learning (DL) can unify both traditional stages into an end-to-end scheme with promising results. In this chapter, we show the most relevant steganalysis techniques using statistical, ML, and DL methods on digital media.