A secure image permutation–substitution framework based on chaos and compressive sensing

Existing secure image acquisition works based on compressive sensing, viewing compressive sensing–based imaging system as a symmetric cryptosystem, can only achieve asymptotic spherical security denoting that the ciphertext only leaks information about the energy of plaintext. Thereby, compressive sensing–based secure image acquisition systems usually work in some scenarios of friendly attack, such as brute-force attack for the entire key space and ciphertext-only attack for the complete plaintext. In this article, a novel permutation–confusion strategy using chaos and compressive sensing is put forward to protect the privacy-crucial images from more threatening malicious attacks, including known-plaintext attack and chosen-plaintext attack. Security guarantee is provided by one-time random projection of compressive sensing, sample-level permutation, and bit-level substitution. It is noteworthy that double confidentiality is embedded in compressive sensing–based image acquisition process using the key-related measurement matrix and sparsifying basis and the randomicity is improved. Simulation experiments and security analyses suggest that the proposed framework can achieve a high-level security and the performance of image reconstruction can be distinctly improved.


Introduction
With the rapid development of multimedia technology, digital image has become an increasingly prevalent information carrier over the Internet. With the awakening of awareness that data are the treasure, personal privacy information contained in digital image is attracting more and more attention from malicious attackers. Image security is of great importance in both military and commercial fields. Fortunately, a large number of image-oriented privacy-preserving techniques are constantly emerging, for instance, image encryption, 1 image digital signature, 2 secret image sharing, and image information hiding. 3 Image encryption is the most straightforward way.
In essence, all kinds of image encryption techniques aim at achieving both diffusion and confusion effects, which were identified as the two essential properties of a secure cipher by Shannon. 4 Diffusion means the statistics redundancy of the plaintext is dissipated into the statistics of ciphertext, while confusion aims to break the connection between the ciphertext and the key. For an encryption algorithm with good diffusion, the ciphertext will change absolutely in a pseudo-random manner when even one bit in the plain domain is changed. Confusion makes it difficult for attackers to reveal the key, though they hold numerous plaintextciphertext pairs generated by the same key. One of the simplest methods to achieve both diffusion and confusion is a permutation-substitution network. 5 In general, permutation is to scramble the position of pixel or bit and substitution is to change pixel values in the field of image encryption. [6][7][8] Expect for the above image encryption techniques, some pioneering works 9,10 tried to embed confidentiality feature in the process of compressed sampling. Compressed sampling (also called compressive sensing, CS), [11][12][13] a merging sampling theory, is able to sample and compress simultaneously without the limit of Shannon-Nyquist sampling theorem and exactly reconstructs the original signal from very few measurements (samples) as long as the measurement matrix obeys some energy-preserving conditions. [14][15][16] When CS is regarded as a symmetric cipher, the original signal, the reconstructed signal, and the measurement matrix are considered as the plaintext, the ciphertext, and the key, respectively. Accordingly, sampling and reconstruction process denote encryption and decryption process, respectively. From this point of view, CS has the potential of integrating sampling, compression, and encryption in the only one physical layer at a minimal cost. CS-based cipher works under a simple fact that it is computationally infeasible to recover the plaintext from the intercepted ciphertext without the complete knowledge of the corresponding measurement matrix. To avoid transmitting a large-scale measurement matrix, chaotic systems were suggested to produce the measurement matrix and the initial parameters (seed values) controlling a chaotic sequence were viewed as the shared key between the encoder and the decoder. 17,18 It should be noted that CS-based cryptosystem is generally required to work under one-time pad, that is, the same measurement matrix is used only once. For reusing the key of CS-based cipher, Zhang et al. 19 proposed a CS-based product cipher by embedding randomness in both measurement matrix and sparsifying basis. In view of the linearity of CS projection, CSbased cipher can achieve computational secrecy rather than information-theoretical secrecy. According to a latest research, 20 attackers could not reveal any valuable information but the energy about the original signal from the noise-like measurements when the measurement matrix consists of the independent and identically distributed (i.i.d.) Gaussian entries. In light of this observation, it was suggested that normalizing all measurements and regarding the normalized measurements as the ciphertext would get a perfect version of CS-based cipher. The fact that CS-based cipher can achieve the asymptotic spherical security was defined by Cambareri et al. 21 Recently, applying CS-based cipher to image encryption has attracted more and more attentions. Under the CS-based cipher model, CS-based imaging system can not only reduce the redundancy of image data but also provide some privacy-preserving properties to some extents. Chai et al. 22 combined CS and steganography to generate a visually secure image. Zhang et al. 23 proposed a scalable image encryption framework at different compression ratios, in which block CS works together with a Sobel edge detector and cascade chaotic maps. George and Pattathil 24 produced the measurement matrix by utilizing multiple chaotic maps. Moreover, they proposed a method of constructing a random measurement matrix based on the linear feedback shift register (LFSR). 25 Zhou et al. 26 created a key-dependent measurement matrix using the circulant matrices and controlling the original row vectors of the circulant matrices with logistic map. Besides, they considered the partial Hadamard matrix as the measurement matrix controlled by chaotic map and the acquired measurements are further scrambled. 27 Liu et al. 28 combined CS and chaos to compress, fuse, and encrypt images simultaneously, in which the random measurement matrix is constructed from a logistic sequence and the acquired measurements are further fused by an adaptive-weighted fusion rule. For the purpose of alleviating the burden of information and communication systems, Zhou et al. 27 proposed a simultaneous image encryption-compression scheme based on two-dimensional (2D) CS and hyperchaos.
In this article, we focus on how to design an imageoriented CS-based product cipher with the help of chaos. Considering that CS-based cipher is a lightweight cipher, we give a supplementary way to improve the security. Finally, a novel image sampling-compression-encryption framework is proposed for secure image acquisition, in which a high-level confidentiality is achieved through two approaches. One is that both the measurement matrix and sparsifying basis controlled by a same chaotic system are regarded as the shared secrets between the encoder and the decoder. Specifically, the elements of measurement matrix come from a chaotic sequence controlled by tent map, and the column vectors of a selected sparsifying basis are randomly permuted by leveraging the same tent map. Even having figured out either measurement matrix or sparsifying basis, attackers are still unable to reconstruct the original signal from CS measurements successfully. The other is that the captured data are further encrypted to breakout the linearity constraint of CSbased cipher by performing the sample-level permutation and the bit-level XOR, which are also manipulated by tent map. Obviously, diffusion and confusion effects can be easily achieved due to good pseudo randomness. In essence, the proposed imaging system is a two-layer cryptosystem consisting of CS-based cipher and chaosbased encryption.
The rest of this article is organized as follows. In section ''Related works,'' we briefly introduce some related knowledge about edge computing and CS-based cryptosystem and then point out some existing challenges in this field. In section ''The proposed framework,'' a double-secrecy encryption framework based on CS and chaos is proposed. Simulation experiments and analysis are given in section ''Experimental results and security analyses.'' Finally, conclusion is drawn in section ''Conclusion.''

Related works
In this section, we introduce some fundamental knowledge about CS theory, chaos theory, and how to embed cryptographic feature in CS.

A brief review of CS
CS is an alternative to Shannon-Nyquist sampling when a signal is sparse or compressible. We denote a one-dimensional (1D) signal to be sampled as a column vector x = (x 1 , x 2 , . . . , x N ) T . Assume that a transform basis C = ½c 1 , c 2 , . . . , c N with size N 3 N can provide a sparse representation for a 1D signal where the column vector s contains only K non-zero entries. Thereby, x is said to be K-sparse in the sparsifying basis C. Fortunately, most natural images can be sparsely or compressively represented in a certain domain through some linear transforms, such as discrete wavelet transform (DWT) and discrete cosine transform (DCT). Note that a compressible signal can be transformed to be sparse after an overwhelming number of small coefficients in a certain domain are made to zeros. In the CS framework, sampling and compression for a 1D signal x can be achieved simultaneously by a linear projection where F is a dimensionality-reducing measurement matrix with size M 3 N (K\M = N ) and y 2 i M is the sensed measurement vector. Evidently, equation (2) does not have a unique solution. A significant achievement is that very few measurements contain all acquired information of recovering x exactly, since the measurement matrix F obeys some energy-preserving guarantees. [14][15][16] In view of the sparsity, an instinctive method of reconstruction is to resolve the l 0 optimization problem s = arg min s s k k 0 s:t: y = Fs wheres andX denote the approximate versions because of the lossy compression of CS. Unfortunately, equation (3) is a NP-hard problem. 29 Equation (3) can be relaxed as the l 1 optimization problem s = arg s min s k k 1 s:t: y = Fs An important precondition is that measurement matrix F must satisfy restricted isometry property (RIP). 14 There are some typical matrices satisfying RIP, including Gaussian and Bernoulli measurement matrix with M ø O(K= log (N =K)) rows. 30 holds for any one K-sparse vector v.

Chaotic encryption
In dynamical systems, chaos is extremely sensitive to initial conditions. Although the systems are deterministic, the sequences produced through iterating chaotic map show excellent pseudo randomness. Applying chaotic theory, the practice of cryptography has attracted much interest in the past 30 years, which utilizes the chaotic sequence to encrypt the plaintext. In view of its good randomness and large key space, thousands of chaotic encryption algorithms have been proposed, all of which take aim at a good diffusionconfusion effect by concealing the connection between ciphertext and plaintext as well as the used chaotic sequence. A common chaotic behavior is the tent map, which is defined as where i = f0, 1, 2, :::g and the parameters fz 0 , mg 2 (0, 1). Unlike the logistic map, the probability density function of tent map approximately follows the uniform distribution. Its great pseudo-random performance and high sensitivity to initial parameters make it almost impossible to find the real key by the exhaustive search to key space or guessing without a clue.

CS-based cipher
When viewed as a cryptosystem at the same time, CSbased information acquisition system can provide a certain level of confidentiality guarantee while compressively sampling. As shown in Figure 1, in the common framework of CS-based cipher, the original signal x, the sensed measurements y, and the measurement matrix F is viewed as the plaintext, the ciphertext, and the key, respectively. In addition, F needs to be updated frequently, even in a one-time setup, resulting from that it is not hard to figure out the fixed F from the sufficient plaintext-ciphertext pairs.
To avoid transmitting the large-scale F, a practical solution is that the pseudo-random number generator (PRNG) is used to construct F in the sense player and the initial parameters (seed values) controlling PRNG are viewed as the key known by only the encoder and the decoder. Recently, using chaotic systems to construct measurement matrix has drawn much attention. Yu et al. 17 employed a logistic map to construct F and proved that such a chaotic matrix may satisfy RIP with high probability. Frunzete et al. 18 adopted tent map instead, which shows a higher strength of resisting statistical attacks in CS-based secure applications. Secure key exchange can be realized through two approaches. One is that the key is generated from the local agreed password. The other is that the key is secretly shared between the encoder and the decoder with the help of key-exchange protocols.
To avoid updating the key of CS-based cipher frequently, Zhang et al. 19 provided a bi-level protected CS framework based on RIP-less theory, 16,31 in which the measurement matrix is a key-controlled RIP matrix, C is a key-related sparsifying basis, and FC À1 is a structural non-RIP matrix. To say the least, attackers have the potential of figuring out the non-RIP matrix FC À1 rather than reconstructing x through knownplaintext attack (KPA) 32 and chosen-plaintext attack (CPA). Therefore, anyone intending to recover the plaintext x from the ciphertext y must hold the real keys of both F and C.

The proposed framework
In this section, a double-secrecy encryption approach based on CS and chaos and an additional chaotic encryption are introduced in detail. Moreover, a novel sampling-compression-encryption framework is proposed to realize secure compressed imaging.
Embedding double secrecy in one-time projection of CS As mentioned previously, one-time random projection of CS can perform image sampling, compression, and encryption simultaneously. Almost all CS-based cryptosystems work under a common framework, in which one-time measurement matrix F generated by a keycontrolled PRNG is viewed as the shared secret between the encoder and the decoder. Generally speaking, the security of the plaintext x mainly depends on the size of the key space and the credibility of the key management. The size of the key space is usually large enough to resist the brute-force attack when a chaotic sequence is used to construct F. The credibility of key management depends on the trust in the key holders, the key-exchange protocol, and the physical security. To improve the security of CS-based imaging system, we propose a dual-image encryption scheme without additional hardware cost.
Most of existing CS-based ciphers are based on the fact that the decoding process is doomed to failure when there is no information about F. In other words, embedding cryptographic feature in the process of CS decoding rests only on the measurement matrix and sparsifying basis C is generally considered to be public. According to a latest study, 19 C can be transformed into an equivalent form C 0 through random permutation.
Definition 2. C and C 0 are two equivalent sparsifying bases (i.e.Cs = C 0 s = x, s where P is a random permutation matrix. Such a random permutation matrix can be generated using the tent map by the following steps: Step 1. Initialize the permutation matrix P to an identity matrix with size N 3 N.
Step 2. Iterate the tent map with the key k C = ft 1 0 , m 1 g to generate a chaotic sequence fC 1 Step 4. Divide the acquired integer sequence Step 5. Exchange the C 1, 1 i th column (or row) and the C 1, 2 i th column (or row) in P.
Obviously, random permutation matrix P is directly related with the key k C of a chaotic system. It means that the permuted sparsifying basis CP would be endowed with the cryptographic feature to some extent when k C is shared only between the encoder and the decoder.
Tent map has been proved to be suitable for constructing measurement matrix F. Accordingly, a certain level of confidentiality provided by chaos can be embedded in process of CS decoding. Such a measurement matrix F can be generated using the tent map by the following steps: Step 1. Initialize the measurement matrix to a null matrix F with size M 3 N .
Step 2. Iterate the tent map with the key k F = ft 2 0 , m 2 g to generate a chaotic sequence fC 2 Step 4. Construct F column-by-column using the regularized sequence fC 2 i g MN i = 1 , as follows . . .
where the scalar ffiffiffiffiffiffiffiffiffi ffi 2=M p is for normalization. It is not hard to come to a conclusion that transmitting the two keys, k F and k C , secretly by a secure auxiliary channel or key-exchange protocol can provide such a CS-based cipher with double confidentiality.

Chaotic permutation-substitution encryption
As mentioned previously, CS-based cipher can only provide a computational guarantee of secrecy and the energy of the plaintext x can be revealed through the cryptanalysis on the ciphertext y. Such security constriction is rooted in the linearity of CS. To improve the secrecy of the acquired data, some chaos-based strong encryption methods, including sample-level permutation and bit-level XOR, are used after CS-based secure compressed imaging. Note that CS measurement y is composed of M real values, so it should be quantized in advance. Here, we use a standard uniform quantizer Q(Á) with levels 2 8 to transform y into a binary format Permutation is to scramble the position of CS measurements, which is done by putting them in a random order. A random index sequence fC 3 i g M i = 1 can be generated using tent map with the succeeding steps: Step 1. Initialize a flag sequence fF i g M i = 1 and an index sequence fC 3 i g M i = 1 to 0, and set the key k p = fz 3 0 , m 3 g and i = j = 1.
Step 3. If F i =0, set C 3 i =h, F i =1, and i=i + 1. Therefore, a scrambled measurement vector y 0 can be generated by the following transformation where f p (Á) is a permutation function. Equation (9) means to reorder the sensed data according to the random indexes. Substitution is further considered to change the statistical property. For a good confusion effect, substitution operation is done by performing the bitwise XOR operation between y 0 and a chaotic sequence. Such a chaotic sequence fC 4 i g M i = 1 can be generated using tent map by the following steps: Step 1. Initialize a sequence fC 4 i g M i = 1 to 0 and set the key k s = fz 4 0 , m 4 g.
Step 2. Iterate the tent map to generate a chaotic sequence fC 4 i g M i = 1 2 (0, 1).
Step 3. Calculate fC 4 The chaotic sequence fC 4 i g M i = 1 controlled by the key k s can be used to encrypt y 0 according to the following transformation where the acquired y 00 is regarded as the ciphertext to be transmitted through the public channels.

The proposed sampling-compression-encryption framework
The proposed CS-based double-secrecy cipher can provide a certain level of security guarantee, which can be implemented in the sense player of a low-cost information acquisition system. When there is a requirement for high confidentiality, some additional strong encryption methods need to be employed. Here, as shown in Figure 2, a sampling-compression-encryption framework is further proposed for secure image acquisition, whose security is guaranteed by the built-in CS-based cipher and the chaotic encryption. Assume that the signal to be sampled is a gray image X with size N 3 N and the corresponding sparse coefficient matrix is S 2 i N 3 N , that is, S = CXC T , where C is a fixed orthogonal DWT matrix. The random permutation matrix P 2 i N 3 N and measurement matrix F 2 i N 3 N are constructed from two chaotic sequences fC 1 i g 2N i = 1 and fC 2 i g MN i = 1 controlled by the keys k C and k F . According to the above research, X = C T SC = (PC) T S 0 (PC), S k k 0 = S 0 k k 0 . The proposed double-secrecy encryption (encoding) process can be implemented by the following one-time projection where Y 2 i M 3 N is the acquired measurements matrix. Obviously, the above process is controlled by two chaotic behaviors, resulting in double confidentiality. When Y could not meet a high requirement of confidentiality, the proposed chaotic encryption based on confusion-diffusion mechanism needs to be further employed. Before it starts, Y must be vectorized and quantized, that is, y b = Q(Vec(Y )). Two chaotic sequences fC 3 i g M i = 1 and fC 4 i g M i = 1 controlled by the keys k p and k s are used to implement the following permutation-substitution transformation Obviously, the whole sampling-compressionencryption process X ! y 00 is controlled by the four keys fk C , k F , k p , k s g and is composed of CS-based lightweight encryption and chaotic encryption. Therefore, theoretically speaking, the proposed secure image acquisition framework can achieve a high-level confidentiality.

Experimental results and security analyses
In this section, several simulations are implemented by MATLAB 2016b in Window 7 to demonstrate the high security of the proposed framework. A standard 512 3 512 Lena image is considered as the test image, which can be sparsely represented by 2D discrete Figure 2. The proposed framework.
wavelet transformation (DWT2). Without loss of generality, the compression ratio is set to 0.5 and the orthogonal matching pursuit (OMP) algorithm is utilized to reconstruct the original image. Based on the tent map, the keys k C = f0:45, 0:60g and k F = f0:50, 0:55g are used to construct the permutation matrix P and the measurement matrix F, respectively, and the keys k p = f0:55, 0:50g and k s = f0:60, 0:45g are used to perform the permutation operation f p and the substitution operation f s , respectively. For a more intuitive result, the ciphertexts are still expressed in the image format. The original image, the CS-based encrypted image, the finally encrypted image, and the decrypted image are shown in Figure 3

Histogram analysis
In the field of image processing, the histogram of an image normally refers to a histogram of the pixel values. The histogram of image is an important statistical characteristic, which is usually considered as an approximation of the density function. Generally speaking, an ideal image encryption algorithm has the ability to realize that the encrypted image would not reveal any statistical feature of the original image.
The histogram of the original image, the encrypted image, and the decrypted image are shown in Figure 4. Obviously, the pixel values of the finally encrypted image well distribute in the interval [0, 255], which stands for an excellent encrypted result. However, the pixel values of the CS-based encrypted image approximately follow Guassian distribution, which is not accidental, resulting from a fact that CS measurements follows sub-Guassian distribution. It can be therefore said that the proposed privacy-preserving framework can conceal the statistical characteristic of the plaintext well.

Correlation analysis
After encryption process is performed, an obvious space correlation for the neighboring pixels in a natural image always appears, which is a kind of meaningful information in some privacy-preserving scenarios. Hence, it is very crucial for an image encryption  algorithm to avoid leaking the correlation. There is a mathematical definition of the correlation coefficient, Cor, used to assess the correlation between two adjacent pixels, which is expressed as follows where x i and y i are the two randomly selected pixel values of test image X, m is the selected pixel pairs, and E(Á) means to compute the mean value.
In our experiments, 2000 pairs of adjacent pixels are selected randomly from the original image, the CSbased encrypted image, and the finally encrypted image in the horizontal, vertical, and diagonal directions. As shown in Table 1, any two adjacent pixels in the original images are tightly correlated with each other but the correlation coefficients of the finally encrypted image are close to 0 in all directions. A more intuitive experimental result is shown in Figure 5. In general, CS-based encryption does well in only the vertical and diagonal directions but does not work in the horizontal direction. And the correlation is completely broken up after chaotic encryption is performed.

Key sensitivity analysis
In the proposed encryption framework, four chaotic sequences related with four keys fk C , k F , k p , k s g are generated by utilizing the tent map. k C and k F controlling measurement matrix and sparsifying basis, respectively, are used to finish the CS-based sampling, compression, and encryption. k p and k s are used to finish the proposed chaotic encryption. Theoretically, anyone who would like to recover the original image (plaintext) X must hold all the above keys at the same time. What is known to us is that chaotic system is extremely sensitive to the input initial parameters (viewed as a key). Here, it will demonstrate how difficult for an attacker to get these real keys without any prior knowledge.
In our experiments, a seemingly tiny change 10 -15 takes place in every key, and then these changed key are used to recover the original image. Note that only one key is changed in every experiment. As shown in Figure 6(a)-(c), the decrypted images can not reveal any meaningful information once one key is false. Figure 6(d) denotes that the decryption (reconstruction) algorithm cannot work completely after the bitwise XOR operation is performed. Here, we emphasize a conclusion that the proposed image encryption scheme can not only flexibly meet the different levels of security requirement but also be considered as a multi-party management system by distributing the four keys to several authorized individuals.

Attack analysis
Brute-force attack. Brute-force attack means that an attacker tries to guess the possible keys and uses them for decryption. In the proposed framework, a chaotic system is used to perform CS-based sampling-compression-encryption, sample-level permutation, and bitlevel substitution operations. The above key-sensitivity test intuitively demonstrates that the whole key space is large enough to make brute-force attack infeasible, which is therefore said to be computational secrecy.
Ciphertext-only attack. Ciphertext-only attack (COA) means that an attacker tries to get some valuable information about the plaintext only from the ciphertext. The security of the proposed framework depends on two kinds of privacy-preserving operations, including CS-based encryption and chaotic permutationsubstitution encryption. In view of the linearity of CS projection, an interesting security feature of the CSbased cryptosystem under COA is the asymptotic spherical secrecy, meaning that any two different plaintexts with the same power keep approximately indistinguishable from their ciphertext. Such a property is attributed to the energy-preserving guarantee of RIP. In the proposed framework, the chaos-based bitwise XOR operation (substitution) can break the energy-preserving property such that attackers would not reveal any valuable information through COA.
KPA and CPA. As for the standalone CS-based cryptosystem, at least N independent plaintext-ciphertext pairs (512 3 512 pairs for 512 3 512 Lena image) are   needed to figure out the used measurement matrix.
Only when the measurement matrix is reused multiple times the attackers are likely to reveal the plaintext from the intercepted ciphertext through CS reconstruction algorithm. But in the proposed framework, the final ciphertext approximately completely loses the feature of measurements such that there is no energypreserving guarantee. Besides, the keys controlling F and C would be frequently updated. Hence, attackers are not able to get both the used F as well as C and the energy information of the plaintext, meaning a failure under KPA or CPA scenario.

Comparison
In this part, the superiority of the proposed framework is demonstrated by comparing it with some existing works. First, we present some potential attack types in this field, including attack type I: brute-force attack to the whole key space or guess at possible measurement matrices, attack type II: reveal the energy information of plaintext by COA, and attack type III: calculate the used measurement matrix by KPA or CPA. Then, we analyzed their resistance to the above attack types. Finally, their security levels are given in Table 2. CSbased sampling-compression-encryption framework, in which single 18 or multiple 24 chaotic map(s) is used to construct measurement matrix, is a lightweight cipher. The bi-level protected CS framework 19 is a prominent work but still could not get rid of the linear relationship between plaintext and ciphertext. In Zhou et al. 26 and Liu et al., 28 some additional privacy-preserving methods are integrated with CS-based cipher to enhance the confidentiality, but their confusion-diffusion effects are still not outstanding. In the proposed framework, not only is CS-based cipher embedded with double secrecy but also sample-level permutation and bit-level substitution (XOR operation) are used to guarantee a great confusion-diffusion effect.

Conclusion
In this article, it is pointed out that CS-based image acquisition systems have the potential to perform sampling, compression, and encryption simultaneously. Besides, how to design the CS-based cipher is deeply studied from two feasible view points, that is, the keydependent measurement matrix and the key-related sparsifying basis. Moreover, we employ a chaotic system to realize a CS-based cipher with double confidentiality. Considering that CS measurements cannot provide an adequate guarantee of confidentiality, an intrinsic weakness of CS-based cipher, an additional encryption approach based on chaos, and diffusionconfusion mechanism have been proposed. Combining the CS-based double-secrecy cipher with the permutation-substitution encryption, we proposed a novel sampling-compression-encryption framework for secure image acquisition. Based on the simulation experiments and analyses, it is therefore said that the proposed framework achieves a high-level security.

Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.  19 Hard Hard Easy Low George and Pattathil 24 Very hard Easy Easy Low Zhou et al. 26 Hard Easy Hard High Liu et al. 28 Hard Hard Hard High The proposed framework Very hard Very hard Very hard Very high