Hybrid Image Compression-Encryption Scheme Based on Multilayer Stacked Autoencoder and Logistic Map

Neetu Gupta,Ritu Vijay

1 Computer Science&Engineering Department,Banasthali Vidyapith,Banasthali,India.

2 Department of Electronics Engineering,Banasthali Vidyapith,Banasthali,India.

Abstract: Secure transmission of images over a communication channel, with limited data transfer capacity, possesses compression and encryption schemes.A deep learning based hybrid image compression-encryption scheme is proposed by combining stacked auto-encoder with the logistic map.The proposed structure of stacked autoencoder has seven multiple layers, and back propagation algorithm is intended to extend vector portrayal of information into lower vector space.The randomly generated key is used to set initial conditions and control parameters of logistic map.Subsequently, compressed image is encrypted by substituting and scrambling of pixel sequences using key stream sequences generated from logistic map.The proposed algorithms are experimentally tested over five standard grayscale images.Compression and encryption efficiency of proposed algorithms are evaluated and analyzed based on peak signal to noise ratio(PSNR),mean square error(MSE),structural similarity index metrics(SSIM)and statistical,differential, entropy analysis respectively.Simulation results show that proposed algorithms provide high quality reconstructed images with excellent levels of security during transmission..

Keywords: compression-encryption; stacked autoencoder; chaotic system; back propagation algorithm;logistic map

I.INTRODUCTION

To store and transmit the information in the form of large and high-quality digital images, compression schemes are of key concern.The principle reason for image compression is to represent and transmit the original large size image with minimal bytes and to reproduce the image with excellent quality at the receiving end.The information contained in an image is constant but different representations of an image lead to variations in the amount of data stored in the image.So during representation with larger amounts of data,some data is irrelevant or redundant.Suitable features can improve the performance of image recognition.Over the past years,features of images were always specified manually that depend on the designer’s prior knowledge,and the number of specified features were very limited.Deep learning models can learn automatically the unlimited number of unpredictable features of image and these unpredictable features can also be used for image compression.Deep learning based compression algorithms have achieved tremendous efficiency for lossy as well as lossless compression techniques[1].

Since images have high redundancy, heavy data capacity, adjacent pixels with strong correlation so existing techniques like DWT, DCT, Huffman coding and fractal algorithms are not suitable for reconstruction of good quality and highly secure images[2].Feature extraction method is required for optimization of image processing [3].In this proposed research,multiple levels of features will be extracted through stacked auto encoder(SAE)model.SAE has multilayer structure that map the input data of large vector representation into a lesser vector dimension.This mapped output converts input data into dense form[4].As a result, SAE can be used for compression of image.

Figure 1.Compression-encryption process.

Data encryption is a process by which the content of data is scrambled.By this, data will be unreadable and incomprehensible during transmission[5].A chaotic system based encryption method is proposed which is sensitive and dependent on initial conditions for achieving high security.In chaotic algorithm, if the chaotic parameters and cryptographic keys are mapped symmetrically then it is difficult to track down the yields without knowing the initial parameters[3].

Confusion and diffusion are the two important aspects to encrypt the image using chaos theory.In diffusion, the pixel values are changed by a certain specific process.In recent years, many research schemes are introduced for confusion and diffusion processes like DNA sequence operation[6,7],matrix semi tensor[8],bit level diffusion[9],two point diffusion[10,11]and perceptron model[12].A piecewise linear maps and one time pseudo random key are used to provide the encryption to image data[13].Scrambling mapping can be generated using piecewise linear chaotic map.Bit level permutation is performed by scrambling matrix [9].The initial conditions of chaotic map should not be stationary to achieve robustness against different attacks.The initial conditions can be vary by using DNA sequence generation[6].

If compression techniques are employed before encryption processes then it reduces the computational time of encryption, possibility of decoding the compressed-encrypted image and cryptanalysis misuse.

In proposed research, a hybrid model to accomplish compression-encryption to the images is depicted in figure 1.Based on SAE,a multi-layer model is constructed to compress the image.An image is put into the first layer and original image is reproduced through the outputs of different level of layers.If any arbitrary layer of proposed model is having its corresponding output smaller than the size of input image then subjective layer will represented as compression stage.The compressed image is further encrypted using chaotic logistic map.This model can be used in tasks that have certain requirements for limited bandwidth and security.The attributes of proposed research are summarized below

1.The plain image is compressed using proposed 7-layers SAE.The compression efficiency of this proposed compression methodology is analyzed based on the compressed image size,PSNR,MSE and SSIM.

2.The compressed image is encrypted through proposed logistic map based encryption methodology.Pixel values of compressed image are confused by pseudorandom sequences generated by logistic map.Bits of pixels are scrambled first in columns then in rows to provide randomness to the encrypted image to achieve more security from attacks.

3.The encryption efficiency of compressedencrypted (CE) image is analyzed based on statistical, differential, key sensitivity, entropy analysis and by evaluating robustness against attacks.

4.The reconstructed image quality is analyzed by calculating PSNR,MSE and SSIM between original image and decrypted-decompressed image.

5.The computational times are calculated for compression-encryption as well as decryptiondecompression processes.

II.RELATED WORK

P.Ramasamyet al.[14] presented enhanced logistic tent map algorithm to generate the key.Colour image is divided into four quadrants which are further divided into four sub quadrants.Pixel values in blocks are scrambled using modified Zigzag transformation.Chaos technique provides randomness to the cipher text.H.Panet al.[15] have studied and analyse the security features of double logistic chaotic map algorithm.Y.Luoet al.[16] have introduced two dimensional baker chaotic algorithm to improve the limited range and security features of a single chaotic map.Pixel positions are scrambled and values are diffuse using new generated chaotic sequence.The improved chaotic map provides randomness and unpredictability to the generated key so that it can enhance the security concern of the transmitted image.To solve key management issue and resist against different attacks,S.Zhuet al.[17]suggested a technique for encryption using SHA-256 and Chaos.Hash value is generated using SHA-256.In the extent of cryptography S.S.Askaret al.[18]presented an encryption technique using two dimensional chaotic economic and logistic map.The security concerns against different attacks are improved by large key space.By combining the characteristics of wavelet transform and chaotic mapping,H.Gaoet al.[19] have presented image compression encryption scheme.Suggested scheme reduces memory consumption, time cost and increases the quality of reconstructed image.Y.Wanget al.[20]fuse the permutation and diffusion to accelerate encryption process and to reduce the scanning efforts.To generate pseudorandom numbers,spatiotemporal chaos is applied.To shorten the key length in measurement matrix, N.Zhouet al.[21] suggested a hybrid algorithm for compression encryption of image in which plain image is segmented into four blocks then scrambling on pixel values of different blocks are introduced.To enhance security during transmission, an encryption algorithm is proposed by G Ye.et al.[22]in which chaotic map and information entropy based encryption algorithm is presented.X.Liuet al.[23] have performed image compression encryption and fusion simultaneously using chaos and compressive sensing technology to process and transmit fewer amounts of data with security.After applying key controlled measurement matrix to original image, fused image is again encrypted using fractional Fourier transform.The proposed scheme also improves the efficiency of key distribution.S.Suriet al.[24] have proposed an evolutionary algorithm to optimize the encryption problems.DNA and Coupled map lattice combination is used to encrypt the image.For better masking of original image multi objective genetic algorithm is proposed.

Z.Chenget al.[25]have used rate distortion less function to trained convolution encoder.Quantization and entropy encoder are applied to generate codes.To exploit statistical dependency of sparse representation L.Polaniaet al.[26]have addressed deep learning architecture with compressive sensing using restricted Boltzmann machine model.S.Haniset al.[27] have suggested a technique for compression-encryption of two images simultaneously through convolution and logistic map.To provide strength against attacks diffusion is performed using discrete nonlinear system cellular automata technique.Huet al.[3] have proposed 5-layers stacked autoencoder and logistic map to compress and encrypt the colour images respectively.The suggested scheme uses two hidden layers at encoder side which limits the compression ratio.In this scheme the initial parametersx0and r are calculated using sigmoid function by which a random sequence is generated which is XOR with compressed image at bit level.This scheme does not show randomness as well as robustness against different attacks.X Chaiet al.[10, 11, 28] have presented compression and encryption techniques by combining the 2D compressive sensing and chaos theory.In Ref.[28], To compress plain image high dimensional measurement matrices are produced and optimized by kronecker product (KP) and singular value decomposition (SVD) respectively.6D hyper chaotic model is utilized for chaotic sequence generation.In Ref.[10],chaotic sequences are generated for confusion and diffusion processes using 4D hyper chaotic and logistic tent systems.In Ref[11],two level of encryption(bit level and pixel level)is used.Bit level as well as pixel level shuffling in Measurement value matrix is performed for confusion and diffusion encryption process.

Figure 2.(a)RBM model(b)Autoencoder model.

X.Wanget al.[8,29—31]have presented numerous ways of efficient image encryption.In Ref.[29]parallel diffusion process is performed through TIDBD method and permutation by combining the sorting and cyclic shift.In Ref.[29] real secret key is generated by filtering the pseudorandom keys through Boolean network.These secret keys are used as initial condition of chaos sequence.In Ref.[8]plain image is scrambled by scrambling the three random positions.Two rounds of diffusion processes are used.In first round Boolean matrix is saved as image and in second round matrix semitensor product is used to generate encrypted image.In Ref.[31] a global pixel diffusion with two chaotic sequences are introduced to improve the security features.Based on the literature review, different encryption and compression techniques have some limitations.In image compression techniques during the learning process many parameters need to set manually that provides limited size of compressed image with poor reconstructed image quality during decompression.During compression and encryption using compressive sensing, linear estimation may change the pixel values and melding of scrambling operation eliminates the adjacent pixel coefficients.Therefore, it may raise the correlation coefficients of encrypted image that provides lower randomness and may cause to assault the encrypted image.In DNA sequence based encryption techniques,extraction of DNA sequences and DNA computing may be easier by combining it with chaos theory but the decryption process is not feasible and also the security features are limited.The confusion process using bit permutation and chaotic map uses the pixel shuffling based on pixel positions.Their randomness is limited and therefore it also limits the security against attacks.The analysis of related work emphasizes the need of a compression algorithm which provides minimum size compressed image and faithful reconstruction of original image during decompression process.A deep learning based SAE model,having three different stages of hidden layers,is proposed to compress the original image.Back propagation technique is used to train the proposed model.In this proposed research, compression is performed before encryption to protect encrypted image against unethical attacks.In this encryption algorithm to encrypt the compressed image every time a different random key is generated to initialize the initial condition parameters for the logistic map so it provides resistance against known plaintext attack (KPA) and chosen plaintext attacks (CPA).The confusion and diffusion processes are performed at bit level and pixel level both to achieve double random encryption approach.

III.PRELIMINARY KNOWLEDGE

3.1 Stacked Autoencoder(SAE)

Figure 3.Logistic map bifurcation and blank window[32].

The auto encoder(AE)is a nonlinear structure used for feature extraction of an image at multiple levels.Two Restricted Boltzmann Machine(RBM)models as shown in figure 2a connected back to back as shown in figure 2b, forms an unsupervised AE model having a single hidden layer.The stacked formation of several AE’s forms a stacked autoencoder model.Previous AE output will be the input to the next AE.Since stack autoencoder is a multi hidden layers model so during designing a stacked autoencoder three stages of different layers are designed.First and third layers are input and output layer respectively and several hidden layers are designed between them.Several AE’s encoding sections are put together one by one and their decoding sections are put together in reverse order.Layer at each level comprises of various nodes.The quantity of nodes at any concealed layer is less than the nodes of corresponding input layers.The SAE structure can extract various levels of features and each level of features represents a compressed image[3].

3.2 Back Propagation Process

Let a multilayer neural network has n+1 inputs represented asx0,x1,...,xnand a bias add on the first hidden layer nodes.Lety0,y1,...,ymare the expected outputs and real outputs areLetL1is the input layer andLnlis the output layer.Here biasesb(1),b(2),...,b(l)are set at a level of+1.The input weighted sum at nodeiinlth layer is represented as

Whereis weight related with the association betweenjth node in (l −1)th layer andith node inlth layer.

Activation value ofith node inlth layer is represented asnd is given by

Sigmoid activation function f is given by

The output error of autoencoder is represented as

For the output layerl=nlset

For other layer i.e.l=nl−1,nl−2,...

By using partial derivatives

3.3 Training Process of SAE

Single layer model of autoencoder can directly trained to find weight by greedy training method.But since stack autoencoder has several hidden layers and if each layer trained individually,it will take large computational time.To reduce the computational time, the training process of stacked autoencoder is performed by two processes i.e.Pre and fine training processes.The output of last hidden layer and the data sequence is used to trained the complete model.In multilayer model,the initial weights directly influence the accuracy and efficiency of the structure.If initial weight is very large then it arises difficulty in locating local minimum.On the other hand,if the initial weight is very small then the gradient of the first layer will also be very less.Pre-training process quantifies the isolated features and gives the suitable initial weights.These weights are utilized in fine training process.The back propagation technique is used for pre training and fine training processes.

?

3.4 Logistic Map

The Logistic map is a one dimensional discrete time map which has unexpected degree of complexity.The logistic map equation is given by-

Here ifi =0 thenx0is called initial value of logistic map and it lies between 0 and 1.λis a varying parameter which lies in the range of 0 and 4.

Ifλ>1, we have a single fix point atx∗=1−1/λloses stability at a critical valueλ=3 and 2-cycle period emerges.Ifλis increased beyondλ =3.449,then 2- cycle period also loses stability and 4-cycle period occurs.Asλapproaches towardsλ=3.54409,4-cycle period also will not be stable and 8-cycle period emerges and this process continued further.Atλ=3.569946,a chaos appears in the system and this stage refers as period of infinity,as shown in figure 3.

Algorithm 2.Image encryption algorithm.

Step 1.A binary sequence of 256 bit session key is randomly generated usingK=round(rand(1,256)), herek=k1k2.

Step 2.Randomly generated 256 bit binary session key is mapped to generate initial parameters(x0, y0)and system parameter r.

Step 3.A succession of real numbers is generated by emphasizing logistic map using generated parameters in step 2 and stored in two dimensional array of size[MN ∗2]and then reshaped.

Number of iterations is decided byiter =ceil(log2(M × N))/log2(4)).HereM × Nis the dimension of input image.

Step 4.Encryption system comprises of confusion and scrambling process.During confusing process reshaped array is divided into blocks for processing and then pixels in image are substituted using given equation where pi represents pixel values of compressed image and xi is pseudorandom sequence value.

Step 5.Scrambling process is done using pseudorandom sequence obtained from step 3 and confused matrix generated from step 4.The pixel values bits are shuffled within column and then within row

[v, Epix]=sort(R(:,:, 1), 1);

for i=1: size(R,1)

C0(:,i)=P(Epix(:, i),í);

Step 6.Decryption is performed by operating above steps in reverse order.Scrambling process is followed by substitution process.

Figure 4.Proposed model of stacked autoencoder.

IV.PROPOSED ALGORITHMS

In the proposed model, three hidden layersh1,h2,h3are used to design the SAE model.In the first stage, SAE is having input layerx, one hidden layerh1and output layerHere input is transferred in to low or high dimensional codeh1by the encoder sigmoid functionfθand thenh1reconstructs the input data x through decoder sigmoid functiongθ.The error is minimized by using back propagation (BP) algorithm and reconstruction error is given by

A second hidden layerh2with output stageis stacked in existing model of AE as shown in figure 4.The hidden layerh2and its output stageare combined with the h1 and form a new AE.This new AE obtain set of parameters using BP algorithm.

By removing the last layera no.of subsequent layers can be stacked in this AE model.In this proposed model,three stages of hidden layers are used so that model may not get complex.First two stages of this structure are called pre training process.

In the last stage, an output layer is stack with the model and initializes the parametersw4andb4between third hidden layerh3and the output layer,which construct the SAE-BP neural network.The simultaneous training of all the weightswiand biasbi,i =1,2,...,lusing BP algorithm is called fine tuning process.

Figure 5. Proposed hybrid compression-encryption process.

The image compression and image encryption process are shown in Algorithm 1 and Algorithm 2 respectively,and the Proposed hybrid compressionencryption process shown in figure 5.

Figure 6. Representation of input test images,corresponding CE images and reconstructed images in Row1,Row2,Row3 respectively.

V.PARAMETRIC EVALUATION AND ANALYSIS

In this section different results are depicted by implementing the proposed algorithms for compression and encryption of an image.The feasibility of these proposed algorithms is validated by implementation on five test images labelled as Baboon,Barbara, Boat, Lina and Pepper.All five test images are of grayscale type,.BMP format and 512×512 size.MATLAB 2018a is used for simulating the proposed algorithms.

Figure 7.Histograms representation of test images.

5.1 Image Compression Efficiency Evaluation Metrics

In the proposed research PSNR, MSE and SSIM are measured and analysed for two different scenarios as discussed below:

• When plain image is only compressed through proposed stack autoencoder based compression algorithm.

• When plain image is compressed-encrypted through proposed stack autoencoder compression and logistic map chaos theory based encryption algorithms.

The size of compressed image received from SAE model is represented in Table 1 and Table 2 when all input images are having size of 257KB.The PSNR values under proposed algorithms for original and compressed images as well as original and reconstructed (decompressed-decrypted) images are shown in Table 1.The PSNR for original and reconstructed image is in between 80.64dB to 88.37dB for all test images whereas for original and compressed images it lies in between 76.30dB and 83.99dB.From analysis of these values, it is observed that PSNR for decompressed-decrypted images are higher than the PSNR of compressed images.It can also be concluded that if only compressed image is transmitted then it contains less PSNR in comparison with if CEimage transmitted.

Table 1.Comparison of PSNR values(in dB).

Table 2.Comparison of MSE values.

Table 3.Comparison of SSIM values.

Table 4.Comparison of variance values.

Table 5.Comparison of correlation coefficients.

To achieve high accuracy and faithful reconstruction of transmitted image at receiver side, MSE should be minimum.The MSE between input and compressed image for the proposed algorithms lies in the range of 0.00025 to 0.0015.For the reconstructed image MSE lies in between 0.00014 to 0.0013 with respect to original image.The results of MSE during execution of proposed algorithms are compared with other existing methodologies of image compression and encryption and it can be viewed that proposed algorithms provide excellent results for all five test images in order to receive minimum value of MSE.

The structural similarity index should approach to unity magnitude in order to define exact similarity between two images.In the proposed research the structural similarity index for original and compressed image lies in between 0.8558 to 0.9145 for all test images and for original image and reconstructed image (decompressed-decrypted) it lies in range of 0.9924 to 0.9986.It is observed that similarity index of input image and reconstructed image is higher than the input and compressed image.The comparative analysis shown in Table 3 results that proposed technique provides higher SSIM than other research schemes.

Figure 8.Correlation distribution of plain image,compressed-encrypted image,reconstructed image.

Table 6.Comparison of NPCR&UACI.

Table 7.Comparison of entropy.

5.2 Image Encryption Efficiency Evaluation Metrics

5.2.1 Statistical Analysis

(i)Histogram Analysis

The experimental analysis of proposed algorithms is carried out on 8 bit grayscale images have 28=256 different intensities.Original input image has sharp variation in distribution of pixels with intensity, so sharp rise and falls occur in histogram of original image.Since encrypted images are coded with binary sequences and have uniform distribution of the pixels with intensity so flat histograms occur for encrypted images.

The histogram analysis shows that histograms for original input image are comparable with histograms of reconstructed images which indicate that images are reconstructed effectively.The uniformity of histogram is analyzed using variance analysis.The variance and uniformity of histograms are inversely proportional to each other [40].Table 4 shows the variance value for original images and their corresponding CE images.Original images have large value of variance therefore does not show any uniformity but the corresponding CE images have low variance in the range of 5451 to 5463 which shows the higher uniformity.Therefore,the proposed technique is capable to provide more security.

Figure 9.(a)Original plain image(b)Encrypted image using key k (c) Encrypted image using key k’ (d) Decrypted image using different decryption key.

(ii)Correlation Coefficient Analysis

The correlation between adjacent pixels can be calculated and analysed with the help of correlation coefficient analysis.Since, the plain image contains key information so adjacent pixel values should be strongly correlated and similar and corresponding correlation coefficients must attain an ideal value approaching towards unity.Encrypted image shows randomness among pixels so adjacent pixels should not be related to each other and be least correlated.The correlation coefficient must approach towards zero value in ideal situation for encrypted image.

Table 5 represents the correlation coefficients of plain as well as CE images for the proposed algorithms.The correlation coefficients for all five test images are in range of 0.726 to 0.9844,which reflects strong correlation among pixels for plain images.The CE versions of five test images are having correlation coefficients in range of−0.00031 to 0.00068,which indicates minimum correlation among pixels.

It can be viewed from figure 8 that correlation distribution of plain images and corresponding reconstructed images are same which authenticate the efficiency of the proposed CE algorithms

5.2.2 Differential Analysis (Plain text sensitivity analysis)

Figure 10. (a) Complete black plain image (b) encrypted image of complete black (c) histogram of complete black encrypted image (d) complete white plain image (e) encrypted image of complete white(f)histogram of complete white encrypted image.

Plaintext sensitivity is analysed by observing the significant variations in encrypted images with only one bit change in plain image[9].Differential analysis shows plaintext sensitivity.NPCR and UACI are the parameters which employ the relationship between two encrypted images and plain image.NPCR provides the pixel value difference between two encrypted images and UACI gives average changed intensity difference of modified encrypted image and plain image.NPCR and UACI should have values more than 99% and 33% respectively for efficient and secure transmission of image data[29].Table 6 represents the NPCR and UACI values of proposed encryption algorithm for all five test images.The experimental analysis shows that proposed algorithm has NPCR and UACI more than 99%and 33%respectively.

5.2.3 Entropy Analysis

In the proposed research pixel intensity of transmitting image is represented by 256 levels (0-255).To represent these 256 levels initially 8 bits are required.The measurement of entropy for plain and CE images are given in Table 7.All five test images in plain form gives entropy in the range of 7.13 to 7.63 while CE form of all test images have entropy in 7.992 to 7.993.

Table 8. Entropies and correlation coefficient values of the plain,encrypted images of complete white and complete black images.

Table 9.Computational time(in Seconds)performance analysis.

5.2.4 Key sensitivity analysis

Key sensitivity shows significant changes in the output by making single bit change in key[44].In proposed technique, initial condition parameters are generated from randomly generated 256 bit key.Key sensitivity is analysed in two ways:(i)same plain image is encrypted using two slightly different keys.(ii)Different key is used to decrypt the encrypted image.Letkandk′are the two keys, differ with one bit, used to encrypt the plain image then two different encrypted images are compared to show the resistance against brute force attack.The generated original keykand one bit modified keyk′are given as:

k=9ab123ec51de01a123bc67ef436de870123def89 ab34cd126734abdcedf56af7

k′=8ab123ec51de01a123bc67ef436de870123def89 ab34cd126734abdcedf56af7

In this proposed algorithm the difference between two encrypted images, as shown in figure 9(b) and 9(c),is calculated based on NPCR and that is found as 99.75%.The decrypted image from different key is shown in figure 9(d)and it is completely different from original image.It proves that it is extremely difficult to crack the original image.

5.2.5 Known/chosen plaintext attack analysis

Chosen-cipher text, chosen-plaintext, knownplaintext and cipher text-only are four classical types of cryptanalysis attacks.Out of which most influential attack is chosen-plaintext attack in this cryptanalysis method, attackers choose their own plain image and generate corresponding encrypted image and try to get the encryption information.If a cryptosystem provides security against chosenplaintext attack then security against three other attacks is also achieved [45].Complete white or complete black images are used by attackers so that permutation process becomes invalid [46][46].To show that proposed algorithm provides resistance against chosen- plain text attack, two encrypted images are obtained from complete black and complete white plain images using proposed encryption technique.From figure 10,it is observed that histograms of both encrypted image is uniform so it does not provide any special information to the attacker.Results in Table 8 show higher entropy values for encrypted images and less correlation values among pixels in encrypted images.So it is observed that the proposed technique provides resistance against known/chosen plaintext attacks.

5.3 Computational Time

Efficiency of the proposed algorithms is indexed with the time taken to execute compressionencryption and decryption-decompression processes for an image when compression is performed before the encryption process.Algorithm efficiency was tested on a system havingi7 processor with 1TB hard disk and 16 GB RAM and Window 10 operating system.MATLAB 2018 is used to simulate the algorithms and 100 rounds are performed to calculate the average execution time.Table 9 presents the computational time performance analysis.

VI.CONCLUSION

From the analysis of compression efficiency parameters,it is concluded that proposed compression algorithm provides high PSNR (>80dB), low MSE (>0.00088),High SSIM(>99%)in compare with other existing compression algorithms.From the analysis of encryption efficiency evaluation parameters,it is also concluded that proposed encryption algorithm provides the secure transmission of image since it provides high NPCR(>99%)and high UACI(>33%).It is also concluded from the results that image can be effectively,efficiently and securely transmitted over a communication channel if simultaneous compression and encryption algorithms are performed.Histogram analysis, entropy analysis and correlation coefficient analysis guarantees the better quality of reconstructed image by decompression and decryption process of the proposed algorithm at receiver end.The presented work in this research can be further extended and analyzed over the colour images of different sizes.The proposed compression-encryption scheme can be further extended by combining the stacked autoencoder compression method with multi objective and optimized encryption techniques.