Hybrid Fuzzy-Enhanced CNN Steganalysis With Optimized Feature Learning For Detecting Low Payload Steganography In Spatial Domain Images

Croix, Ntivuguruzwa Jean De La Croix (2025) Hybrid Fuzzy-Enhanced CNN Steganalysis With Optimized Feature Learning For Detecting Low Payload Steganography In Spatial Domain Images. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

Text 7025221024-Doctoral.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Progress in Deep Learning (DL), particularly Convolutional Neural Networks (CNNs), has significantly improved the accuracy of identifying images with possible hidden data. The data hiding paradigm involves a method and an artistic endeavour to conceal confidential data within ordinary digital content, such as digital images. Departing from the data hiding paradigm, known as steganography, steganalysis is counter-art aiming to identify whether hidden data is present within digital media. Over the last few decades, steganalysis and steganography have maintained a symbiotic relationship, with each being interchangeably employed and mutually fostering the advancement of the other. In any steganographic strategy in images, the primary objective is to optimize the imperceptibility of the secret information in the content of an image, thereby preserving the integrity of the original cover image. The high imperceptibility of a stego image makes it difficult, and mainly impossible, to discern the presence of the secret bits. Therefore, the primary endeavour of steganalysis revolves around discovering whether an image is an original cover or a stego image.
Existing steganalysis methods have yielded promising results with conventional Machine Learning (ML) techniques; however, the introduction of Convolutional Neural Networks (CNNs), a deep learning paradigm, has achieved better performance over the previously proposed ML-based techniques. Contemporary CNN models for the steganalysis of digital images employ several approaches, including transfer learning and data augmentation, to enhance the accuracy of detecting and locating steganographic payloads within images. Nevertheless, many state-of-the-art approaches rely on stacking numerous convolutional layers to expand the local receptive fields, which showed a common drawback to effectively detecting and locating the possible hidden data with low payload steganography. The challenge of the existing steganalysis systems poses a significant problem in information security to optimize the security of data transmission. This makes this research a considerable contribution by improving the detection and location of low payload steganography.
This thesis focuses on detecting and locating low payload steganographic data in spatial domain images by addressing critical limitations in existing steganalysis methods. It introduces a hybrid fuzzy-enhanced CNN steganalysis framework developed through three key strategies. First, feature learning is enhanced by optimizing CNN design using a combination of small-sized kernels, depthwise separable convolutions, and a spatial pyramid pooling (SPP) module to capture rich multiscale spatial features. Second, dataset representability is improved by applying fuzzy logic-based image preprocessing paradigms that adaptively enhance feature clarity based on image complexity. Third, a fuzzy-enhanced CNN model is developed by integrating a fuzzy reasoning layer within the network to handle feature uncertainties through fuzzy set representations.
Experimental results across BOSSBase 1.01 and BOWS 2 datasets show significant improvements over state-of-the-art methods, with detection accuracy gains ranging from 4.6% to 26.3%. Localization models achieved up to 98.2% F1 scores and up to 56.0% improvement in altered pixel localization for specific payloads. Ablation studies confirm the critical role of fuzzy logic in improving detection and localization. Recommendations for future work include scaling to real-time and color images, as well as integrating attention mechanisms to enhance feature focus.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Information security, fuzzy logic, national security, securing network infrastructure, steganalysis, steganography
Subjects:	Q Science > QA Mathematics > QA76.9.A25 Computer security. Digital forensic. Data encryption (Computer science)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55001-(S3) PhD Thesis (Comp Science)
Depositing User:	Ntivuguruzwa Jean De La Croix
Date Deposited:	31 Jul 2025 05:57
Last Modified:	31 Jul 2025 05:57
URI:	http://repository.its.ac.id/id/eprint/124836

Actions (login required)

View Item