This study proposes an improved video steganography method that combines data embedding based on the Discrete Cosine Transform (DCT) with a deep learning-driven scene change detection mechanism. Unlike traditional rule-based detection, the proposed method utilizes a lightweight convolutional neural network (CNN) to accurately identify scene transitions, thereby improving carrier frame selection across diverse video content. Additionally, an adaptive thresholding strategy based on local image statistics is introduced to strengthen embedding imperceptibility and robustness. The confidential data is securely shuffled using a pseudo-random sequence derived from a user-defined key and embedded into the DCT coefficients of 8 × 8 blocks across the carrier frames. The experimental outcomes confirm that the system indicated possesses superior Peak Signal-to-Noise Ratio (PSNR), better resilience to compression and noise attacks, a higher payload capacity while maintaining video quality. The enhancement of security and robustness of video steganography by utilizing adaptive methods and deep learning is greatly reinforced by what this work has proposed. This method is appropriate for current multimedia applications.