Wasfy B. Mikhael’s research while affiliated with University of Central Florida and other places

In the last couple decades, several successful steganography approaches have been proposed. Least Significant Bit (LSB) Insertion technique has been deployed due to its simplicity in implementation and reasonable payload capacity. The most important design parameter in LSB techniques is the embedding location selection criterion. In this work, LSB insertion technique is proposed which is based on selecting the embedding locations depending on the weights of coefficients in Cosine domain (2D DCT). The cover image is transformed to the Cosine domain (by 2D DCT) and predefined number of coefficients are selected to embed the secret message (which is in the binary form). Those weights are the outputs of an adaptive algorithm that analyses the cover image in two domains (Haar and Cosine). Coefficients, in the Cosine transform domain, with small weights are selected. The proposed approach is tested with samples from the BOSSbase, and a custom-built databases. Two metrics are utilized to show the effectiveness of the technique, namely, Root Mean Squared Error (RMSE), and Peak Signal-to-Noise Ratio (PSNR). In addition, human visual inspection of the result image is also considered. As shown in the results, the proposed approach performs better, in terms of (RMSE, and PSNR) than commonly employed truncation and energy based methods.

Over the past few decades, there have been several successful methods developed for steganography. One popular technique is the insertion method, which is favored for its simplicity and ability to hold a reasonable amount of hidden data. This study introduces an adaptive insertion technique based on the two-dimensional discrete Haar filter (2D DHF). The technique involves transforming the cover image into the wavelet domain using 2D DWT and selecting a predetermined number of coefficients to embed the binary secret message. The selection process is carried out by analyzing the cover image in two non-orthogonal domains: 2D discrete cosine transform and 2D DHF. An adaptive algorithm is employed to minimize the impact on the unrepresented parts of the cover image. The algorithm determines the weights of each coefficient in each domain, and coefficients with low weights are chosen for embedding. To evaluate the effectiveness of the proposed approach, samples from the BOSSbase and custom databases are used. The technique’s performance is measured using three metrics: mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM). Additionally, a visual inspection by humans is conducted to assess the resulting image. The results demonstrate that the proposed approach outperforms recently reported methods in terms of MSE, PSNR, SSIM, and visual quality.

Minimally perturbed adversarial examples were shown to drastically reduce the performance of one-stage classifiers while being imperceptible. This paper investigates the susceptibility of hierarchical classifiers, which use fine and coarse level output categories, to adversarial attacks. We formulate a program that encodes minimax constraints to induce misclassification of the coarse class of a hierarchical classifier (e.g., changing the prediction of a 'monkey' to a 'vehicle' instead of some 'animal'). Subsequently, we develop solutions based on convex relaxations of said program. An algorithm is obtained using the alternating direction method of multipliers with competitive performance in comparison with state-of-the-art solvers. We show the ability of our approach to fool the coarse classification through a set of measures such as the relative loss in coarse classification accuracy and imperceptibility factors. In comparison with perturbations generated for one-stage classifiers, we show that fooling a classifier about the 'big picture' requires higher perturbation levels which results in lower imperceptibility. We also examine the impact of different label groupings on the performance of the proposed attacks. Supplementary information: The online version contains supplementary material available at 10.1007/s00034-022-02226-w.

Convolutional neural network (CNN) has emerged as one of the primary tools for image classification. In particular, deep CNNs are now considered the main tool for this purpose. However, these networks are often large and require computing and storage power that may not be available in very small sensor devices such as IoT (Internet of Things) devices; their training is also time and computing power consuming. As a result, in some applications, reducing the size of input data (images) and the processing network becomes necessary. Such reduction usually comes at the cost of reduced classification performance. In this paper, we consider networks with under 200k learnable parameters, as opposed to millions in deeper networks. We examine how domain transforms can be used for efficient size reduction and improvement of classification accuracy for small networks. We emphasize that finding optimal hyperparameters or network configurations is not our objective in this paper. It is shown that by using transforms such as discrete wavelet transforms (DWT) or discrete cosine transform (DCT), it is possible to efficiently improve the performance of size-reduced networks and inputs. We demonstrate that in most cases, the improvement can be traced to higher entropy of resized input using transforms. While transforms such as DCT allow variable input and network sizes to be utilized, DWT proves to be very effective when significant size reduction is needed (improving the result by up to 5%). It is also shown that input size reduction of up to 75% is possible, without loss of classification accuracy in some cases. We use two datasets of small images, including Fashion MNIST and CIFAR-10, to evaluate the performance of size reduction methods.