Abstract
Deep learning-based image steganalysis has progressed in recent times, with efforts more concerted toward prioritizing detection accuracy over lightweight frameworks. In the context of AI-driven health solutions, ensuring the security and integrity of medical images is imperative. This study introduces a novel approach that leverages the correlation between local image features using a CNN fused Long Short-Term Memory (LSTM) model for enhanced feature extraction. By replacing the fully connected layers of conventional CNN architectures with LSTM, our proposed method prioritizes high-relevance features, making it a viable choice for detecting hidden data within medical and sensitive imaging datasets. The LSTM layers in our hybrid model demonstrate better sensitivity characteristics for ensuring privacy in AI-driven diagnostics and telemedicine. Experiments were conducted on Break Our Steganographic System (BOSS Base 1.01) and Break Our Watermarking System (BOWS) datasets, followed by validation on the ALASKA2 Image Steganalysis dataset. The results confirm that our approach generalizes effectively and would serve as impetus to ensure security and privacy for digital healthcare solutions.