Secure Spread Spectrum Image Steganography Using a CNN-Based Learned Detector

Spread Spectrum Image Steganography (SSIS) represents a promising approach for embedding secret data into a cover image. In conventional methods, a pseudo-noise (PN) sequence functions as a secret key, without which neither message embedding nor data extraction is feasible. However, this secret key presents a potential security risk, as an adversarial service may attempt to uncover it, and if successful, unauthorized access could enable the extraction of the secret data. This study introduces a novel steganography technique inspired by spread spectrum principles. Unlike conventional methods that embed information directly using PN sequences and employ correlator-based detection, the proposed method does not depend on the PNs themselves as the primary carrier. Instead, it constructs structured PN-based patterns and encodes the secret message within these patterns rather than within the raw PNs. As a result, the reliance on PN values is eliminated in the proposed Decoder/Encoder steganographic framework, thereby enhancing the security of proposed approach compared to traditional methods. To extract the secret data, a convolutional neural network (CNN) model is employed to classify the received PN pattern and determine the corresponding pattern class. In fact, while the method reduces dependence on explicit PN sequences, security is transferred to CNN parameters and mask configuration. Experimental results indicate that the proposed CNN-based method is not only competitive with other deep learning-based approaches but also outperforms conventional SSIS techniques under the evaluated attacks. Moreover, as a secondary contribution, the method offers an additional advantage over conventional SSIS techniques by enhancing robustness against some geometric attacks, a well-known limitation of traditional approaches.