Coverless Steganography Based on Low Similarity Feature Selection in DCT Domain

Anotace: Coverless image steganography typically extracts feature sequences from cover images to map information. Once the extracted features have high similarity, it is challenging to construct a complete mapping sequence set, which places a heavy burden on the underlying storage and computation. In order to improve database utilization while increasing the data-hiding capacity, we propose a coverless steganography model based on low-similarity feature selection in the DCT domain. A mapping algorithm is presented based on an 8000-dimensional feature termed CS-DCTR extracted from each image to convert into binary sequences. The high feature dimension leads to a high capacity, ranging from 8 to 25 bits per image. Furthermore, scrambling is employed for feature mapping before building an inverted index tree, considerably enhancing security against steganalysis. Experimental results show that CS-DCTR features exhibit high diversity, averaging 49.3% complete mapping sequences, which indicates lower similarity among CS-DCTR features. The technique also demonstrates resistance to normal operations and benign attacks. The information extraction accuracy rises to 96.7% on average under typical noise attacks. Moreover, our technique achieves excellent performance in terms of hiding capacity, image utilization, and transmission security.