Integrating Hybrid Encryption and Deep Learning for Robust Audio Steganography and Audio Genre Classification

In steganalysis, modern steganographic methods outperform more traditional machine learning approaches like SVM and ECs. To this end, Ntivuguruzwa With a focus on digital picture steganography, De La Croix et al. [1] examine cutting-edge advancements in deep learning-based steganalysis. The growth of these approaches has been analyzed throughout the study, with special emphasis on their superior speed, precision, and efficiency. The outcomes show how deep learning techniques improve steganalysis dramatically, and they provide reliable detection for modern steganographic systems. While secure communication is important for vehicular ad hoc networks, most traditional approaches often fail in this regard. Ansari [2] examine in their study whether the use of picture steganography could improve data security and secrecy in VANETs. For this end, the author presents several steganography techniques including vector embedding, spatial domain, and transform domain and computes the efficiency of such techniques when the losses of visual quality are compensated by ensuring safe data transmission. The results are shown below, and indicate that, even in challenging network environments, picture steganography represents a workable solution for secure communication on VANETs. Patient data security is increasingly significant due to the wide diffusion of telemedicine services and healthcare data exchange. Ghosh et al. [3] conduct research into how efficient the steganography techniques used are for the protection of medical information in e-health networks. The work will discuss many techniques like LSB, PVD, transform domains and the trade-off between security and quality parameters, such as PSNR and SSIM. The study concludes with a premise that the best and safest method for data transmission in telemedicine is a combination of diagonal queue-based steganography with chaotic techniques and Rabin encryption. One of the very important problems to solve in steganography is the problem of hidden embedding in images. Martín et al. [4] investigate whether GANs enhance spatial domain steganalysis techniques when combined with hidden information, showing few changes. With these results, a GAN is shown to be able to avoid the optimal Deep Learning steganalysis attacks when it is trained to modify an image in order to hide a message by use of the LSB algorithm. The results here describe how much more effective and secure this is when it comes to steganography. Upon hiding of the secret information by steganography within texts, its statistical distribution may be changed significantly as its quality decreases and makes it difficult to hide. To address this problem, Sun et al. [5] proposed a graph-to-text generation -controlled topic hiding method. The method uses graph route coding instead of conditional probability while generating knowledge network steganographic writings by configurable subjects. Experimental results compared with the original scheme indicate that this approach greatly improves the hiding of steganographic messages and text quality. Because linguistic steganography is generation-based, this procedure frequently causes the statistical distribution of the original text to be disturbed and consequently lowers the quality of the text and reduces hiding ability. Rustad et al. [6] propose a technique known as graph-to-text generation-informed topic-controlled steganography to overcome the problem. In the absence of conditional probability, the method derives steganographic texts from the knowledge graph such that the topics are controlled and carry secret information as the graph route is coded. Experimental results show that this method enhances text quality and offers the best cover capacity when compared to traditional ones, and it suggests being an effective steganographic tool. One of the main issues is guaranteeing the secrecy of the secret information but still safe over the Internet. To overcome the difficulty involved, Durafe and Patidar [7] have presented here a dependable and blind color picture steganography technique utilizing fractal cover images, SVD, IWT, and DWT. For protection against processing assaults and to preserve picture quality, it very smartly conceals a large amount of hidden image behind a tiny fractal cover image. Yazdanpanah et al. [8] discuss the inefficiency of the steganalysis algorithms currently in use for speech audio files. This paper describes the feature Percent of Equal Adjacent Samples (PEAS) for speech steganalysis with a Gaussian membership function-based classifier to capture 50% of embedded stego occurrences. Li et al. [9] have developed a unique steganalysis technique that can successfully identify MV-based steganography in HEVC. The method entails determining Motion Vector Prediction's (MVP) local optimality and examining how message embedding compromises it. The best rate of MVP is shown by the approach as a steganalysis characteristic. According to experimental results on two datasets, this method provides low-complexity, practical detection without the need for model training, o0utperforming four state-of-the-art strategies. Protecting the ownership of neural networks requires model watermarking, yet model pruning attacks pose a threat. To tackle this problem, Li et al. [10] provide a watermarking approach that prevents pruning for the hidden image steganography auto-encoder. The method includes employing three traditional pruning algorithms to choose the proper model weights and using a DCT-based technique to implant a watermark. The fourth and fifth decimal positions of these weights include the watermark. High robustness is demonstrated by the experimental findings, which also exhibit gains in watermarking capacity and watermark extraction accuracy above 0.9993, even after 40% pruning. In response, Gao [11] provide a convolutional neural network (CNN)-based visual identification technique that combines deep learning with image processing theories. The study evaluates the method's advantages and disadvantages and makes recommendations for improvements. Results from experiments show that the CNN-based method provides excellent resilience and accuracy, which substantially improves automated visual recognition performance. To solve these problems, Sushma et al. [12] present MARVIS, a novel data embedding technique based on the Mellin transform. The method greatly increases detection resistance and data embedding capability. The experimental findings demonstrate that MARVIS outperforms conventional techniques and increases data capacity without affecting cover object integrity, achieving a PSNR of 50-60dB and an SSIM of 0.9998 for embedding 4 bits of data. The swift growth of communication networks and online services has raised questions about how to protect sensitive information while it is being sent. Back-propagation learning is a revolutionary steganography approach proposed by Dhawan et al. [13] to overcome this problem. The technique uses pixel value differences and modification direction to improve the output of least significant bit replacement using a hybrid fuzzy neural network (HFNN). With a PSNR of 61.96 and an MSE of 0.041 for the secret picture, experimental findings show that the technology outperforms conventional methods in achieving great security and outstanding visual quality. Because certain characteristics in deep steganalysis are discarded, training on photos with different content frequently results in lower detection accuracy. In response, Mo et al. [14] provide a picture steganalysis technique that makes use of deep content feature clustering. Deep CNN are used to extract low-frequency content characteristics, which are then grouped for pre-classification, and wavelet transform is used to eliminate high-frequency noise. This method increases consistency between training and testing pictures by up to 4.84% for J-UNIWARD at 0.4 bits per non-zero alternating current discrete cosine transform coefficient and 1.39% at 0.2 bits per non-zero alternating current in terms of detection accuracy. Compared to picture and audio steganography, text-based steganography has received less research, despite its cheap bandwidth overhead. In order to close this gap, Alqahtany et al. [15] have proposed a multi-layered, dynamic, and resilient steganography technique that uses Arabic diacritic elements to obfuscate critical data. For security and performance, the technique uses text, encryption schemes, and imagery. It has been evaluated under many scenarios. Since the method turns out to be an excellent solution for text-based steganography, it is resistant to known attacks and preserves the integrity of carrier text apparently unaltered. Such neural network algorithms can detect classic audio steganography because it typically alters cover audio's properties. Li et al. [16] introduced an innovative coverless audio steganography model that utilizes the WaveGAN framework in direct synthesis of stego-audio, in order not to disturb any existing covers. The extractor can recover the hidden audio successfully since it is accompanied by resolution blocks. The method comes with full semantic transfer and excellent resilience. Due to the growth of online data transmission, protecting these data becomes increasingly important. For the purpose of enhancing cover concealment and resilience, Pilania et al. [17] recommend an steganography approach based upon JPEG compression combined with Integer Wavelet Transform. The method improves the qualities of an imperceptibility and resistance using the properties native to video cover files and JPEG compression. It explores and analyses the level of resistance of the approach against attacks based on image processing through PSNR, MSE, SSIM and CC. The results show high concealment with low complexity in both concealing and retrieval processes along with great security. Steganography is the art of hiding messages inside media such as pictures in a way that the hidden data is hard to be detected. Hammad et al. [18] have undertaken the task of steganalysis, which is finding hidden messages in symmetric pictures despite having identical characteristics. They present a classification method that combines Gaussian discriminant analysis (GDA) and naïve Bayes (NB) classifiers with texture features including segmentation-based fractal texture analysis (SFTA), local binary pattern (LBP), and gray-level co-occurrence matrix (GLCM). Steganography is the art of hiding secret messages within other digital information such as audio, images, and videos in such a way that nobody detects it. To mitigate against this, Shehab and Alhaddad [19] outline steganalysis techniques that try to detect and recover hidden signals as well as measure the resistance of the employed steganography methods. In an attempt to describe this better, their work identifies general concepts as well as classifying the methodologies used in steganalysis techniques. It discusses in detail the current practices in steganalysis of various media, discusses the current limitations, and future directions, and thus, gives valuable insights into the development of the field. Internet-of-Things devices with network connectivity have introduced the cybersecurity threats with which safe data transfer requires. Towards this, Djebbar [20] have proposed a low-overhead, noise-resistant solution for steganography over Internet-of-Things settings. Their technique tests the accuracy of concealed data against noise and different wireless technologies using a variety of modulation and coding techniques. With low bit error rates, great undetectability, and cheap complexity, the technique efficiently conceals large payloads in audio signals, making it suitable for IoT device deployment. 3D pictures are becoming a new area of interest for steganography, which has progressed to incorporate a variety of cover files. In their comprehensive overview of 3D steganography research during the previous fifteen years, Tanwar et al. [21] classify the work according to evaluation parameters, security aspects, and algorithms. Additionally, they examine pertinent steganalysis methods for 3D steganography. Their analysis provides a thorough summary of the state-of-the-art techniques and obstacles while highlighting the advancements in 3D steganography and suggesting topics for further study. In their review of Arabic text steganography, Thabit et al. [22] note this problem of protecting sensitive information in public channels and mention this technology's potential, since text has low bandwidth, and it is hard to search for hidden messages in an ocean of content in the internet. To evaluate several Arabic text steganography methods, their study classifies them based on four parameters: robustness, security, invisibility, and capacity. One of the most significant barriers to effective covert information transfer in voice signal steganography is synchronization, which Wojtun and Piotrowski [23] addressed. They then come up with four new methods of synchronization, three of which directly take the acoustic signal, while the fourth probes the structure of the decoded steganographic data stream. Evaluations use both objective and subjective quality ratings, providing a measure of the success of these strategies. Alhaddad et al. [24] discuss audio steganography in MP3 files and specifically focus on the problem of securely transferring private information across IoT networks. Al-Rekaby et al. [25] proposed a method by combining multi-level chaotic maps with Least Significant Bit (LSB) steganography and Advanced Encryption Standard (AES) encryption, this study improves security approach for text transmission. Also Abuali et al. [26] proposed method that uniquely combines Dynamic Least Significant Bit (DLSB) steganography with Wavelet Obtained Weights (WOW) steganographic algorithms, forging a sophisticated and adaptable system for secret data embedding. AbdAl-Hameed et al. [27] introduced an image steganography approach, leveraging double density dual tree wavelet transform (DDDT-DWT), designed to enhance capacity while preserving optimal quality, it produced 97.92% accuracy at 128 kbps compression. Problem in detection of steganography in low bit rate audio streams inactive VoIP settings [28]. It identifies and introduces a new steganalysis technique that uses poker test statistics to encode bits of parameters influencing the frequency of subsequences and will appear symmetric after the steganographic embedding.

4.1 Data preparation, feature extraction and labelling

This research is in deep learning audio classification. In this step, we aggregate and preprocess a wide spread of files for diverse genres of audio. We then extract the Mel-frequency cepstral coefficients from these files as a way of obtaining the essential properties of the audio. This is because MFCCs are powerful tools of audio analysis as they are highly effective in representing audio spectrum. Extraction involves loading the audio files into segments and computing the MFCC features for each segment. This structured dataset is the output of MFCC vectors.

4.2 Hybrid encryption and QIM

The hybrid encryption technique extends the strengths of RSA and AES to protect hidden messages in audio files. It uses RSA as an asymmetric encryption to encrypt a sensitive message such that it can only be read by the recipient with the corresponding private key. This key is encoded in the audio with the Least Significant Bit (LSB) method, which means that the LSB of each audio sample is replaced by the bit of the AES key. In the extraction phase, one can obtain from the audio file the AES key and the RSA-enciphered message and thus decrypt it in two successive steps, the first decryption step is given to be provided by AES decrypting the ciphertext, and the second decryption step by using the private key by RSA to yield the original message. This hybrid approach is sure to deliver sound, secure transmission with strong security.

Introducing adaptive audio steganography is through utilization of the Quantization Index Modulation (QIM), Least Significant Bit (LSB), and Spread Spectrum (SS) techniques, aimed at enabling strong message embedding into audio files of the GTZAN dataset. The idea is to embed secret information inside the audio files of the GTZAN dataset, as it is done using a multi-level approach for improving security and imperceptibility. Considering QIM and its application in framing-level energy thresholds, the embedding procedure is optimized for audio regions with a higher energy and results in minimum perceptual distortion. In this case, the interference with the original signal is reduced as it uses the least significant bits of audio samples. Last but definitely not the least, the Spread Spectrum method is a new variation that spreads the covert message throughout the signal by a pseudorandom sequence that has relatively hard time in being extracted by an unauthorized party.

Applicability of these methods across the different genres within the GTZAN dataset makes them immune and applicable to the given audio characteristics. The comparison of plots between waveform and frequency spectrum unveiled the invisibility of the steganographic techniques and the high level of Spread Spectrum robustness in terms of detection along with wide distribution of information. In Figure 2, the GTZAN dataset encryption and the AES and RSA keys were generated. A hybrid approach provides advanced status for audio steganography by offering improved security, imperceptibility, and adaptability of the employed audio genres.

2.jpg

Figure 2. GTZAN dataset encryption and generation of keys for AES and RSA

4.3 Model evaluation and performance metrics deep learning comparative analysis

Audio data is classified as encrypted and steganographic, in this work based on local, spatial, and temporal patterns of information captured by deep learning models intrinsic to audio signals. CNNs are applied on the spatial patterns through convolutional layers to extract rhythmical and spectral features in audio signals. Pooling could be used within CNNs to compress the dimensionality in order to focus the model's ability on the most crucial aspects of audio to process large datasets efficiently while preserving critical information.

3.jpg

Figure 3. Audio steganography applied quantization index modulation on GTZAN dataset

RNNs, GRUs, and the LSTM model are designed to handle the temporal dependencies within sequential audio data. RNNs retain the hidden states, meaning that the network is capable of remembering past inputs; therefore, these are specially useful for time series data. GRUs build on top of the traditional architecture of an RNN in terms of adding gates to control information flow that help the model to capture long-term dependencies. In Figure 3, the Audio steganography applied quantization index modulation on GTZAN were shown, LSTMs further improve this ability in a network that, in principle at least, could selectively retain information over longer sequences-a feature quite crucial for classification tasks based on the evolving characteristics of audio data.

The Performance Comparision of CNN, RNN, GRU and LSTM models for different genres as shown in Figure 4, CNN gains consistency in having a very high accuracy of 0.9502 but performs poorly in the terms of precision and recall. In fact, the precision even drops to 0.5714 and 0.5696 in Country and Rock genres respectively. In contrast, RNN has significantly lower accuracy at 0.8224; yet, it still demonstrates the better precision and the value of F1-score for some genres like Blues and Classical at 0.95. The GRU and LSTM demonstrate similar accuracies at both: 0.9444 and 0.9486 respectively. In the case of Classical and Pop, the model has better accuracy and recall but a lower F1-score in Rock.

4.jpg

Figure 4. Comparision methods for various parameters using deep learning techniques

Of all models, it is noticed in the Table 1 that the worst result in overall evaluation is shown by RNN, and, more seriously, in terms of accuracy at 0.8224 compared with other ones. Despite the fact that the model shines in genres, such as Classical, where the F1-score reaches the value of 0.9544, precisions and recall in genres of Rock (0.6748 - precision, 0.7023 - recall), Blues (0.7654 - precision) are worse than other models. This reflects lower RNN's performance on various datasets, probably due to its incapability to capture long dependencies as efficiently as LSTM or GRU.

Table 1. Comprehensive analysis of accuracy, precision, recall, and F1-score for genre classification models

Experimental Evaluation under Noisy Conditions: To assess the robustness of the proposed system, additional experiments were performed by introducing Gaussian noise (20dB SNR), low-bitrate compression (64kbps MP3), and downsampling (22.05kHz) into the GTZAN dataset. The results showed that while CNN achieved the highest accuracy on clean audio (95.02%), its performance dropped more sharply under noisy and compressed conditions (89.47% and 87.12%, respectively) compared to LSTM, which maintained higher resilience with accuracies of 91.25% and 89.63%. This indicates that although CNNs excel in clean environments, recurrent architectures such as LSTM are better suited for real-world deployment where, audio signals are often degraded, thereby demonstrating stronger robustness and adaptability of the proposed hybrid encryption and classification framework.

Encryption Performance Evaluation: The efficiency of the proposed hybrid RSA–AES encryption was compared with AES-only and RSA-only methods on 1 MB audio files. AES-only was the fastest (≈12 ms/MB) but less secure in key management, RSA-only was highly secure but very slow (≈238 ms/MB), while the hybrid method achieved a balanced trade-off (≈16 ms/MB) by combining AES speed with RSA security. This demonstrates that the hybrid scheme preserves strong protection with only a minor performance overhead.