The AutoThing framework [3] presents a secure transaction framework for a diverse self-service things (SST) environment. It demonstrates that traditional SST systems are not resilient against cyber-attacks and introduces an Offline Self-Service Things (OSST), which has higher attack resilience. Adee and Mouratidis [10] propose a four-step security consisting of RSA encryption, LSB steganography, data backup and recovery, and data sharing. Awotunde et al. [18] advocate for employing blockchain technology and multi-level authentication to secure transactions in banking platforms. Luo et al. [19] added blockchain to strengthen the security of steganographically concealed data transmission in covert communication. Jahnavi and Nandini [20] propose combining Steganography, Cryptography, and Blockchain to secure the Banking ecosystem. These aforementioned efforts represent multi-layered frameworks that integrate two or more security mechanisms to elevate protection levels for banking applications. However, limited research [19,20] specifically addresses banking data security through the combined application of steganography and blockchain. The present work distinguishes itself by delivering a practical, implementable solution for self-service banking applications through the strategic integration and adaptation of IoT-Cloud infrastructure, Generative Adversarial Networks (GAN), Steganography, and Blockchain paradigms.

In a self-service banking system, (i) only authorized persons should be permitted into the system, and (ii) the customers’ information/transactions should be completely secure. The main contributions of this paper include:

IoT-Cloud-based secure architecture for self-service Banking

The rest of the paper is organized into the following sections. Section 2 describes the preliminaries of Generative Adversarial Networks, Steganography, and Blockchain, and Section 3 presents the background of security in the Banking domain. A review of recent Literature is presented in Section 4. Section 5 explains the proposed solution methodology. The experimentation and performance analysis are described in Section 6. Conclusion and future work are presented in Section 7.

The Generative Adversarial Network comprises of two neural networks—a Generator and a Discriminator, that are trained simultaneously through an adversarial competing process. These networks engage in continuous opposition: the Generator network generates new data by taking as input random data, and the Discriminator classifies the Generator’s fake data from real data. This adversarial process drives both networks to improve over time. The objective of the GAN program is to achieve a Nash equilibrium, wherein the Generator produces data of such realism that the Discriminator cannot reliably differentiate between generated and genuine samples. Upon reaching this equilibrium state, the Generator has effectively learned to capture the underlying data distribution, producing high-quality synthetic output. GANs are used for a variety of data types, including images, music, and text. GANs have been successful in various tasks such as image generation, style transfer, super-resolution, and even generating realistic human-like text. Common applications of GAN include—facial recognition for contactless entry (DigiYatra), face-speech authenticated access control, emotion identification, rare disease diagnosis, student attention monitoring, and improved astronomical images [21].

In this research work, GAN is employed to produce accurate representations of human faces [22]. The preparatory step for face recognition involves generating clear frontal images of the customer from any-angle face images captured by the IoT camera. The Generator model learns to synthesize plausible frontal facial images while the Discriminator evaluates authenticity by comparing outputs against real customer images from the training dataset. The resultant frontal image is verified with the bank’s customer KYC-database to authenticate the customer to gain access to banking operations. The various components of the GAN face recognition are depicted in Fig. 1.

Steganography takes as input a secret digital message and a cover image, to generate imperceptible image perturbations and adds these perturbations to the cover image to get the steganographic (stego) image [23]. The basic model of image steganography encompasses several steps (Fig. 2).

(i)

Cover Image Selection: The first step is to select an image that disguises the confidential information. It is often chosen so that it appears trustworthy to the viewers.

Blockchain technology is characterized by its decentralized architecture, open-source nature, anonymity (or pseudonymity), and immutability. Blockchain offers immutability, time-stamping, traceability, transparency, non-repudiation, and automation [26,27]. Blockchain technology is being increasingly explored across diverse businesses and sectors, and its specific effects in banking and financial services are examined [28]. The convergence of blockchain and cryptographic technologies has established a transformative commercial paradigm with enhanced security and superior features over real currency. Fundamentally, Blockchain is a decentralized, distributed digital ledger used to record and keep up to date various types of digital transactions in today’s world.

The general architecture of Blockchain is depicted in Fig. 3. The Blockchain distributed ledger mechanism is implemented using multiple blocks connected in a chain-like structure. Transactions are represented by creating new blocks that are connected with each other. Every new block contains the previous block’s hash, and every block has a unique hash; hence, the blockchain is tamper-proof. Each transaction contains the cryptographic hash, transaction date, transaction time, and transaction data. The important aspect of blockchain is that all blocks agree on a network-verified transaction using consensus mechanisms such as Proof-of-Work (PoW), Proof-of-Authority (PoA), Proof-of-Stake (PoS), etc.

The problem is to design a seamless and secure self-service banking system. Every customer who desires to perform any banking operation should first be properly authenticated. His/her data and transactions should be confidential and immutable. The key requirements and proposed solution for the problem is as depicted in Fig. 4.

Step 1: The IoT camera captures the facial image of the customer as he approaches the terminal for commencing his/her transaction. Since the customer may not pose in a perfectly straight orientation, the face angle may be tilted to the left, right or facing upwards or downwards. This IoT camera will be rendered as a standardized frontal image for biometric authentication in Step 2.

Step 2: A highly accurate frontal orientation image of the customer is generated using GAN. This image is matched with the customer KYC image and the customer gets authenticated to access his/her account.

Step 3: The customers data should be protected during transmission in network channels as well as during storage in Cloud. An improved hybrid Curvelet-LSB steganography is used to conceal the confidential data (producing a stego-image). The embedded data can be extracted on the other side to retrieve the original data.

Step 4: Blockchain is used for transactions and storage security. Customer transactions are uploaded to Blockchain, providing time-stamping, traceability, non-repudiation and immutability.

Step 5: Finally, the confidential data is stored in the Cloud permanent storage using Blockchain process.

The proposed solution offers a practicable, secure self-service Banking mechanism. For instance, when a customer comes to a Bank, he seamlessly enters the premises by automated GAN facial authentication; he can go to his smart locker inside the Bank and access it without the need for any escort staff. Similarly, a customer can carry out mobile internet banking or access the ATM with biometric facial authentication. His transactions are secure with Steganography and Blockchain.

A customized GAN for this application is described in the accompanying pseudocode. A Generator and a Discriminator are the two basic elements in this arrangement. The Generator constructs the frontal image that seems like a real image. The Discriminator, on the other hand, tries to tell the difference between the image that came from the Generator and the real image. The training procedure is iterative and spreads over several rounds (epochs). The Discriminator is trained to recognize actual data during each epoch before being trained to recognize fake data produced by the Generator. The Generator is then taught to improve the data it generates, based on the Discriminator’s input. The Generator tries to fool the Discriminator with its created data, while the Discriminator seeks to get better at detecting the Generator’s outputs. In short, GAN is a continual game between the two. The various components of the GAN system (Algorithm 1) are as follows: (i)

Random Input: It is the input to the Generative models. In this case, it is the different angled facial images of a person. Using this, the GAN can be trained to generate a wide range of synthetic data.

The Generator-Discriminator model is updated in batches, specifically with a collection of real samples and a collection of generated samples. An epoch is defined as one pass through the entire training dataset. The discriminator model is updated twice per batch, once with real samples and once with fake samples, which is a best practice as opposed to combining the samples and performing a single update. The front-angle image generated by GAN should be accurate for the authentication of the customer.

For Face Authentication, a normalized size (240 × 320 pixels), frontal orientation proper image of the customer is input to the Discrete Wavelet Transform (DWT) for comparison. DWT decomposes the input face image into LL, LH, HL, and HH subbands (Lower Lows, Lower Highs, Higher Lows, and Higher Highs) to represent the horizontal, vertical, diagonal, and approximation coefficients, respectively. The steps in face recognition (Algorithm 1) are—feature extraction, classification using a GAN model, and comparison of images [44].

Thus, the GAN Biometric Authentication is applied in two phases: first to transform a customer’s any-angled picture into a frontal image, and second to identify the customer.

A hybrid method combining Curvelet transform and LSB techniques is developed to hide the confidential information inside a cover image without affecting the visual quality and size of the cover image. The corresponding stego image containing the secret data can be transmitted without being found out. This method allows the intended receiver to recover secret information without much loss of data. Before applying LSB steganography, the curvelet based denoising step is applied as a preprocessing step to eliminate noise from the cover picture, which is an undesirable by-product that adds false and unnecessary information. This improves the image quality. Edge preservation is improved by the curvelet transform, which also offers superior denoising than the wavelet method. The curvelet transform is a multiscale geometric analysis technique that displays image features at every scale and in different directions. providing better curve handling and overcoming the wavelet shortage problem at higher frequencies.

This work uses the least significant bit (LSB) to conceal data inside discrete curvelet coefficients. As a preprocessing step before steganography, this method uses the curvelet denoising technique to eliminate noise from the cover image. Using the Curvelet Matlab toolbox, it applies a rapid discrete curvelet transform to the cover picture, based on the unequally spaced rapid Fourier Transform (FDCT- USFFT). The LSB method is employed to insert the secret message into the transformed coefficients, and then the inverse fast discrete curvelet transform (IFDCT-USFFT) is used to recover the message and reconstruct the image.

The proposed Steganographic method (Algorithm 2) details the process of embedding secret data into a cover image using a hybrid curvelet transform and least significant bit (LSB) technique.

A cover image of size 255 × 255 is obtained. As a preprocessing step, the curvelet denoising is done on this cover image. It helps to remove noise and make the embedding process smoother. The curvelet transform is then applied to the denoised cover image. This transformation helps in extracting features or enhancing the properties of the image, which makes it more suitable for embedding. After transformation, the cover image is divided into multiple blocks, each of size 8 × 8. The secret image is then embedded into these blocks using the LSB technique. LSB embedding is a popular method for steganography where the least significant bits of an image’s pixels are altered to embed secret data. This embedding continues until all blocks of the cover image have been visited. Once the embedding process is complete, the inverse curvelet transform is applied to convert the transformed image back to its original spatial domain. The final result is the stego image, which is the cover image containing the embedded secret.

Procedure Extraction describes how to extract the secret data from a stego image. It begins by acquiring the stego image, which contains the embedded secret. A curvelet transform is applied to this stego image for feature extraction or to get it in a form where extraction becomes feasible. This transformed image is divided into blocks, similar to the embedding process. For each 8 × 8 block of the transformed stego image, the bits of the embedded secret data are retrieved essentially reversing the LSB embedding process. After retrieving the bits from all blocks, they are rearranged or recompiled to reconstruct the original secret. The final result is the extracted secret message. In essence, the Embedding algorithm hides secret data within a cover image using curvelet transforms and LSB embedding. The Extraction algorithm extracts the hidden secret from the stego image.

Blockchain is a promising technology that has the potential to overcome fundamental issues that have long been hampering the banking and finance sector. By reducing expenses, increasing efficiency, and adhering to legal requirements, blockchain technology enables an organization to become more secure, transparent, decentralized, and efficient. Blockchain technology, generates an unchangeable record of transactions, helps to deter theft and other types of illegal fraud. Traditional banking services have seen significant change thanks to blockchain technology advancements. Decentralisation and security are guaranteed by the blockchain ledger’s distribution throughout the bank’s network nodes. As fewer or no middlemen are involved, transactions are completed more rapidly. Processing transactions takes less time and money when blockchain is used. The immutability of blockchain adds more security against fraud. The various functionalities of Blockchain in IoT-Cloud Banking application are explained in Algorithm 3.

In Banking, customers’ identities should first be obtained and verified through KYC processes to prevent fraud. The verified customers can conduct their banking operations. Blockchain is used to ensure traceability, immutability, and non-repudiation of all customer transactions, and store the resulting data in the Cloud. This process is achieved by the following four steps.

The proposed design uses a centralized cloud infrastructure taking into consideration the particular needs of the bank. Preliminary study assists in estimating costs and creating a design for the centralized structure, making sure it incorporates redundancy, data storage strategies, and strong security measures [47]. The next crucial step is choosing a reliable cloud infrastructure (or cloud service provider) that complies with legal and regulatory norms. To prevent duplication or mistakes from being sent during the transition, data purification is performed before migration to the centralized system. Importantly, rigorous access restrictions and strong security mechanisms are used throughout this process. The infrastructure is frequently checked for regulatory compliance and periodically evaluated to maintain compatibility with the bank’s developing goals and needs, especially given the sensitive nature of the business.

The system used for the experiment is an Intel Core i5/2.7 GHz/16 GB/1 TB/4 GB NVIDIA GeForce RTX 3050 Ti PC with a 1 Gbps internet link. The program is written in Python, running on Java Eclipse Environment with PyTorch and Keras; Matlab is used for simulations and plotting.

The different metrics to validate the performance are as follows: (i)

GAN Facial Authentication: Until recently, most GAN developers fell back on the qualitative evaluation of GAN generators via manual assessment or visual inspection of the images synthesized by a generator model, as there were no objective functions defined for the Generator Model to compare the performance. However, the difference between the generated image vs. the real image gives the quantitative measure of GAN success using Frechet Inception Distance (FID) and Inception Score (IS). Precision, Recall, and Fmeasure assess the accuracy of facial recognition systems.

It is difficult to detect the difference between a stego-image and a cover-image by visual inspection. Hence, the PSNR, MSE, and Histogram metrics are employed to evaluate the success of the steganographic technique. For testing and analysis of the hybrid steganography, a file containing basic information such as the customer’s image, name, account number, and transaction details, which is about 67 KB in size, is taken as a constant secret message (confidential data), in this experiment.

PSNR (Peak Signal-to-Noise Ratio) compares the image quality to that of a reference image. It is calculated by measuring the difference in pixel intensity between two images and determining their maximum value. The higher the PSNR value, the better the match between the two images, which means they are of higher quality. In steganography, PSNR calculates the invisibility of the image, that is, to assess the extent of image degradation caused by embedding the secret within the cover image. It is mathematically expressed as:PSNR=10×log 10×MAX2MSEwhere MAX² is the maximum pixel value of the image (for a grayscale image, it is 255). MSE stands for Mean Squared Error between the original and stego-image. It gives the average value of the pixel-by-pixel difference between the original image and stego-image. MSE is inversely proportional to PSNR.

A histogram is a bar graph where the x-axis represents the range of possible pixel values (e.g., 0 to 255 for grayscale images) and the y-axis represents the frequency count of pixels for each intensity. If there is a significant difference between the histogram of the original image vs. the stego-image, it indicates the presence of hidden content in the stego-image.

The robustness of steganography is measured by Bit Error Rate (BER), i.e., (number of incorrect recovered bits)/(total payload bits). When the receiver extracts the hidden message from the stego file, some of the bits may be incorrect compared to the original secret message. Hence, it is desirable to have a lower BER. The errors may be introduced by the steganographic algorithm when it embeds the secret data into the cover image or get introduced due to noisy communication channels.

Steganography Result Analysis

The confidential message is embedded into different types of cover images, and the imperceptibility is evaluated. Standard publicly available grayscale and color images are chosen as cover images. A constant-size secret message mentioned above is concealed in all the cover images used in this experiment. Table 5 shows the PSNR, MSE, and BER for stego images from 6 different standard cover images (3 grayscale and 3 colored) after embedding the said customer’s secret data. The histogram graphs of different cover images and the corresponding stego images after embedding an unchanging secret message is shown in Fig. 10.

A higher the PSNR value, indicates better the match between the two images, and they are of higher quality. And a lower value of MSE implies that the error is less. Typically, the PSNR for color images is higher than the PSNR for grayscale as the depth of hiding is 24 for color pixels. The PSNR value depends on the size of the secret data being concealed. In this experiment, the secret message is kept constant, and hence the PSNR compares the concealment capacities of different cover images for the hybrid Curvelet-LSB steganography algorithm. It is observed that the PSNR values for many stego images are above 40 dB, which indicates that the steganographic algorithm is very effective in hiding the secret data. In this experiment, the average PSNR is 39.57 dB, and the Standard Deviation is 5.58 dB. Hence, the color images (like Jetplane) handle embedding very well with minimal distortion, while others (like Peppers or Cherry Blossom) degrade more noticeably.

The histogram analysis compares the overall shape, intensities of peaks and troughs, distortions, and shifts in the graphical plots for the original against the stego image. The different grayscale intensities or color values are depicted on the horizontal x-axis, and the frequencies of occurrence of the colors are represented on the vertical y-axis. The input host images can be Grayscale images (8-bit image per pixel) or RGB color images (24-bit image per pixel), which impact the hiding capacity of the cover.

A comparison of cover vs. stego histographs for grayscale images for a secret message of 67 KB is performed. The distortion is very high in the histograms of Peppers’ grayscale for the given small secret message. The distortion of histograms for Cherry Blossom grayscale and Rose grayscale is slightly less than Peppers grayscale (Fig. 10). Among the three grayscale cover images, the histogram analysis shows the least difference for the Rose grayscale. From this histogram analysis, it is observed that opaque, dark images have more similar cover-stego histograms. Hence, using dark and opaque images provides better security.

Similarly, histogram analysis of cover vs. stego histographs for color images for a secret message (67 KB) are performed. The cover vs. stego histograms pairs for each of the color images are similar. The frequency values at all pixels are very similar, implying that the imperceptibility is very good for all color cover images. The hiding capacity of color images is high. The grayscale images show more differences between the histograms of cover and steganographic images, whereas the color images have fewer differences in the histograms of corresponding cover and stego images. This is due to the higher hiding capacity of the color images. Overall, it is observed that the size of the secret to be concealed, the steganographic algorithm, and the hiding capacity of the cover image determine the quality of steganographic imperceptibility.

The robustness of hybrid Curvelet-LSB Steganography can be analyzed from the Bit Error Rate (BER). The grayscale images show higher BER (15%–20%) while the color images show lower BER (7%–9%).

The server authorizes the transaction to proceed after successful authentication. Every transaction, including money transfers, is noted on a blockchain database. An aggregation of banking data is periodically uploaded to the Cloud using Blockchain security. The important aspects to be considered while choosing a Blockchain for an application are—public or private mode of operation, consensus mechanism used, upload time, upload data size and transaction fees.

Blockchain Result Analysis (i)

Upload Time: In the context of Blockchain, the upload time is the total time taken to send the file from a device to the server. It is dependent on the data size, network bandwidth/speed and blocktime (time required to store the data inside each block on the blockchain). In other words, upload time is dependent on the type of consensus algorithm that needs to be executed before a block gets committed. The Ethereum Blockchain platform is used in this experiment, as it supports fees-free off-chain mode, in addition to the public on-chain mode. Fig. 11 shows the upload time in Ethereum for different sizes of data. For on-chain mode a gas price of 44 Gwei was incurred. The upload time is less than 3 mins for 100 KB data while it is slightly more than 3 min for 800 KB data. Besides, the off-chain mode (leveraging IPFS) is around 7% faster than the on-chain mode. Typically, a secure color stego image with embedded secret customer information is about 700 KB, while the grayscale stego image is around 300 KB. An upload time of about 3 mins for 800 KB data size is a nominal time to safely store the stego-image by Ethereum blockchain in the cloud.