ENHANCING COVERLESS IMAGE STEGANOGRAPHY THROUGH OPTIMIZED DEEP LEARNING TECHNIQUES. PhD thesis, King Fahd University of Petroleum and Minerals.

PDF RANA_GREEN_COVER 22-merged_Final.pdf Restricted to Repository staff only until 1 March 2027. Download (8MB)

Arabic Abstract

عرف ال steganography image بانه عملية اخفاء بيانات سرية داخل الصورة بحيث نقوم باخفاء وجود البيانات وليس تشفيرها.تتمثل المشكلة في هذا الحقل بكيفية ايجاد التوازن الصحيح بين ثلاثة عوامل هما الsecurity capacity ,imperceptipility .في هذا البحث قمنا ببناء وحساب ثم التنبؤ بخريطة تسمى ال map ttention باستخدام خوارزميات الذكاء الاصطناعي ثم قمنا بتخصيص عدد البتات لكل بكسل عبر عتبات محددة بحيث تُحسب سعة الإخفاء وتُعدّل الصورة السرية لتتناسب مع الهدف . تعتمد عملية الاخفاء على وضع بتات في أقل البتات )LSB( مع ترتيب مواضع البتات بواسطة تتابع قابل لإعادة التوليد باستخدام HMAC-DRBG ومفتاح permutation) (keyed واختياريًا تبييض البتات )whitening( عبر تدفق — HMAC بالاعتماد على خريطة انتباه مصغّرة داخل الهيدر" التي تضيف إلى الهيدر تمثيلا ًمُكمَّمًا لصورة الانتباه المصغّرة )mini( بحيث يستطيع المستقبِل إعادة بناء خريطة انتباه تقريبية .الهيدر نفسه يحتوي على nonce وطول الرسالة ويُكرَّر كل بت ثلاث مرات لاستخراج أكثر موثوقية عبر تصويت الأغلبية. يتوفر أيضا ًمُولّد إعدادات بسيط قائم على خوارزمية جينية )GA( للبحث عن عتبات وعدد بتات جيدين بناء ًعلى مقاييس جودة .SSIMو PSNR مثل

English Abstract

Image steganography is the practice of concealing secret information within a dig- ital image so that the presence of the message is difficult to detect. The central prob- lem is to balance three competing goals: imperceptibility (minimal visible distortion), payload capacity (how many bits can be hidden), and robustness/security (reliable recovery under noise/processing and resistance to steganalysis) while meeting practi- cal constraints such as a target bits-per-pixel (BPP), etc., with a deterministic and key-controlled embedding. This research introduces a novel steganography system, which combines HMAC-DRBG-based keyed adaptive LSB embedding with Attention map-in-header technology, generated through CNN and GAN models, along with xxv the genetic optimization for adaptive bit allocation (to enhance the embedding pro- cess). Our steganography system starts by generating pixel-level attention maps from cover image edges or texture (using HED, RCF, VGG16, and ResNet50), then con- verts them into multiple threshold-based bit allocation levels, and adjusts the secret data size to achieve a specific bits-per-pixel (BPP) rate. Our system also uses HMAC- DRBG-based keyed permutations to protect the secret image’s data from unautho- rized access by applying stream whitening options, while maintaining deterministic behavior for stego header and data placement. We have also considered a nonce and the stego header payload length with threefold bit repetition (majority voting) to en- hance robust recovery capabilities. Additionally, a mini-attention mechanism reduces a quantized, low-resolution attention map (to fit within the header), so receivers can generate an approximate attention map when edge or texture detectors are available. A simple genetic algorithm is also used to find the best threshold and bit-allocation settings that achieve maximum reconstruction quality (PSNR and SSIM) within speci- fied BPP constraints. Finally, at the sender side, the secret image is embedded within the cover image, guided by our attention map (generated using the deep learning model), to produce the stego image, and at the receiver side, the secret image is ex- tracted from the stego image. In addition to the proposed adaptive image steganog- raphy framework, this research integrates a Counting-Based Secret Sharing (CBSS) mechanism to further strengthen data confidentiality and reliability. CBSS divides a secret key into multiple shares and distributes them among authorized participants, allowing the original secret to be reconstructed only when a predefined threshold num- ber of shares are combined. This ensures secure group-oriented access and prevents xxvi any single participant from recovering the secret independently. The proposed CBSS adopts a block-masking strategy, where each 64-bit block of the secret key is XORed with a predefined sequence of ones, producing randomized, ambiguous shares that enhance confidentiality while minimizing computational cost. Experimental results demonstrate that increasing the block size decreases similarity among shares, thereby improving overall security. This dissertation proposes a novel and integrated steganographic framework that combines content-aware attention maps, keyed adaptive LSB embedding, and genetic optimization. The core novelty lies in bridging coverless feature extraction (edges and textures) with embedding-based steganography through attention-guided bit al- location. In addition, the proposed framework introduces a robust header design and incorporates a Counting-Based Secret Sharing (CBSS) mechanism to enhance reliabil- ity and secure multi-party access. Collectively, these contributions advance the state of the art in secure, adaptive, and resilient image steganography. xxvii