Rohit Agrawal

Rohit Agrawal, Kapil Ahuja, Marc C. Steinbach et al.

We hide grayscale secret images into a grayscale cover image, which is considered to be a challenging steganography problem. Our goal is to develop a steganography scheme with enhanced embedding capacity while preserving the visual quality of the stego-image as well as the extracted secret image, and ensuring that the stego-image is resistant to steganographic attacks. The novel embedding rule of our scheme helps to hide secret image sparse coefficients into the oversampled cover image sparse coefficients in a staggered manner. The stego-image is constructed by using ADMM to solve the LASSO formulation of the underlying minimization problem. Finally, the secret images are extracted from the constructed stego-image using the reverse of our embedding rule. Using these components together, to achieve the above mentioned competing goals, forms our most novel contribution. We term our scheme SABMIS (Sparse Approximation Blind Multi-Image Steganography). We perform extensive experiments on several standard images. By choosing the size of the secret images to be half of the of cover image, we obtain embedding capacities of 2 bpp (bits per pixel), 4 bpp, 6 bpp, and 8 bpp while embedding one, two, three, and four secret images, respectively. Our focus is on hiding multiple secret images. For the case of hiding two and three secret images, our embedding capacities are higher than all the embedding capacities obtained in the literature until now. For the case of hiding four secret images, although our capacity is slightly lower than one work, we do better on the other two goals; a) very little deterioration in the quality of the stego-images and extracted secret images, and b) inherently and designed-to-be resistant to steganographic attacks. Additionally, we demonstrate that SABMIS executes in few minutes, and show its application on two real-life problems.

Rohit Agrawal, Kapil Ahuja

Image steganography plays a vital role in securing secret data by embedding it in the cover images. Usually, these images are communicated in a compressed format. Existing techniques achieve this but have low embedding capacity. Enhancing this capacity causes a deterioration in the visual quality of the stego-image. Hence, our goal here is to enhance the embedding capacity while preserving the visual quality of the stego-image. We also intend to ensure that our scheme is resistant to steganalysis attacks. This paper proposes a Compressed Sensing Image Steganography (CSIS) scheme to achieve our goal while embedding binary data in images. The novelty of our scheme is the combination of three components in attaining the above-listed goals. First, we use compressed sensing to sparsify cover image block-wise, obtain its linear measurements, and then uniquely select permissible measurements. Further, before embedding the secret data, we encrypt it using the Data Encryption Standard (DES) algorithm, and finally, we embed two bits of encrypted data into each permissible measurement. Second, we propose a novel data extraction technique, which is lossless and completely recovers our secret data. Third, for the reconstruction of the stego-image, we use the least absolute shrinkage and selection operator (LASSO) for the resultant optimization problem. We perform experiments on several standard grayscale images and a color image, and evaluate embedding capacity, PSNR value, mean SSIM index, NCC coefficients, and entropy. We achieve 1.53 times more embedding capacity as compared to the most recent scheme. We obtain an average of 37.92 dB PSNR value, and average values close to 1 for both the mean SSIM index and the NCC coefficients, which are considered good. Moreover, the entropy of cover images and their corresponding stego-images are nearly the same.

Rohit Agrawal

Steganography plays a vital role in achieving secret data security by embedding it into cover media. The cover media and the secret data can be text or multimedia, such as images, videos, etc. In this paper, we propose a novel $\ell_1$-minimization and sparse approximation based blind multi-image steganography scheme, termed $\ell_1$SABMIS. By using $\ell_1$SABMIS, multiple secret images can be hidden in a single cover image. In $\ell_1$SABMIS, we sampled cover image into four sub-images, sparsify each sub-image block-wise, and then obtain linear measurements. Next, we obtain DCT (Discrete Cosine Transform) coefficients of the secret images and then embed them into the cover image\textquotesingle s linear measurements. We perform experiments on several standard gray-scale images, and evaluate embedding capacity, PSNR (peak signal-to-noise ratio) value, mean SSIM (structural similarity) index, NCC (normalized cross-correlation) coefficient, NAE (normalized absolute error), and entropy. The value of these assessment metrics indicates that $\ell_1$SABMIS outperforms similar existing steganography schemes. That is, we successfully hide more than two secret images in a single cover image without degrading the cover image significantly. Also, the extracted secret images preserve good visual quality, and $\ell_1$SABMIS is resistant to steganographic attack.

Rohit Agrawal, Thibaut Horel

The families of $f$-divergences (e.g. the Kullback-Leibler divergence) and Integral Probability Metrics (e.g. total variation distance or maximum mean discrepancies) are widely used to quantify the similarity between probability distributions. In this work, we systematically study the relationship between these two families from the perspective of convex duality. Starting from a tight variational representation of the $f$-divergence, we derive a generalization of the moment-generating function, which we show exactly characterizes the best lower bound of the $f$-divergence as a function of a given IPM. Using this characterization, we obtain new bounds while also recovering in a unified manner well-known results, such as Hoeffding's lemma, Pinsker's inequality and its extension to subgaussian functions, and the Hammersley-Chapman-Robbins bound. This characterization also allows us to prove new results on topological properties of the divergence which may be of independent interest.

Rohit Agrawal, Yi-Hsiu Chen, Thibaut Horel et al.

We introduce hardness in relative entropy, a new notion of hardness for search problems which on the one hand is satisfied by all one-way functions and on the other hand implies both next-block pseudoentropy and inaccessible entropy, two forms of computational entropy used in recent constructions of pseudorandom generators and statistically hiding commitment schemes, respectively. Thus, hardness in relative entropy unifies the latter two notions of computational entropy and sheds light on the apparent "duality" between them. Additionally, it yields a more modular and illuminating proof that one-way functions imply next-block inaccessible entropy, similar in structure to the proof that one-way functions imply next-block pseudoentropy (Vadhan and Zheng, STOC '12).

Kartik Hegde, Rohit Agrawal, Yulun Yao et al.

The past several years have seen both an explosion in the use of Convolutional Neural Networks (CNNs) and the design of accelerators to make CNN inference practical. In the architecture community, the lion share of effort has targeted CNN inference for image recognition. The closely related problem of video recognition has received far less attention as an accelerator target. This is surprising, as video recognition is more computationally intensive than image recognition, and video traffic is predicted to be the majority of internet traffic in the coming years. This paper fills the gap between algorithmic and hardware advances for video recognition by providing a design space exploration and flexible architecture for accelerating 3D Convolutional Neural Networks (3D CNNs) - the core kernel in modern video understanding. When compared to (2D) CNNs used for image recognition, efficiently accelerating 3D CNNs poses a significant engineering challenge due to their large (and variable over time) memory footprint and higher dimensionality. To address these challenges, we design a novel accelerator, called Morph, that can adaptively support different spatial and temporal tiling strategies depending on the needs of each layer of each target 3D CNN. We codesign a software infrastructure alongside the Morph hardware to find good-fit parameters to control the hardware. Evaluated on state-of-the-art 3D CNNs, Morph achieves up to 3.4x (2.5x average) reduction in energy consumption and improves performance/watt by up to 5.1x (4x average) compared to a baseline 3D CNN accelerator, with an area overhead of 5%. Morph further achieves a 15.9x average energy reduction on 3D CNNs when compared to Eyeriss.

Kartik Hegde, Jiyong Yu, Rohit Agrawal et al.

Convolutional Neural Networks (CNNs) have begun to permeate all corners of electronic society (from voice recognition to scene generation) due to their high accuracy and machine efficiency per operation. At their core, CNN computations are made up of multi-dimensional dot products between weight and input vectors. This paper studies how weight repetition ---when the same weight occurs multiple times in or across weight vectors--- can be exploited to save energy and improve performance during CNN inference. This generalizes a popular line of work to improve efficiency from CNN weight sparsity, as reducing computation due to repeated zero weights is a special case of reducing computation due to repeated weights. To exploit weight repetition, this paper proposes a new CNN accelerator called the Unique Weight CNN Accelerator (UCNN). UCNN uses weight repetition to reuse CNN sub-computations (e.g., dot products) and to reduce CNN model size when stored in off-chip DRAM ---both of which save energy. UCNN further improves performance by exploiting sparsity in weights. We evaluate UCNN with an accelerator-level cycle and energy model and with an RTL implementation of the UCNN processing element. On three contemporary CNNs, UCNN improves throughput-normalized energy consumption by 1.2x - 4x, relative to a similarly provisioned baseline accelerator that uses Eyeriss-style sparsity optimizations. At the same time, the UCNN processing element adds only 17-24% area overhead relative to the same baseline.