Undetectable capabilities, capacity, and security are the biggest challenges facing information hiding in images. This paper presents methods to address these challenges. It utilizes an advanced deep learning architecture combined with a chaotic, multi-level encryption system. Lorenz and Rossler maps are used as chaotic methods to encrypt the message, which is then segmented to increase its resistance to chaotic analysis. High-level features are extracted using ResNet-50 CNNs, and the image is classified into four regions based on entropy. This results in an optimal payload distribution (16x16 to 2x2). The method employs two techniques: Least Significant Bits Plus Plus (LSB++) for spatial modulation in textured regions and Discrete Cosine Transform (DCT) in smooth regions, which increases capacity and detectability. The results were obtained by evaluating two datasets, BOSSBase and COCO, with an average load of 4.2 bits per pixel, a signal-to-noise ratio (PSNR) higher than 54 dB, a structural similarity index (SSIM) higher than 0.97, and a modulation detection error rate of 89% using SRNet. The proposed method demonstrates significant improvements in modulation capacity and detection resistance compared to current methods.