In this work, we introduce a new method for image encryption that improves on the traditional Rubik cube based encryption model by integrating adaptive key generation using Convolutional Neural Networks (CNN) and Pseudo Random Number Generator (PRNG). Chaotic masking with Rubik cube based pixel permutation was employed in while effective, these masking employ static or random key vectors that are independent of the image content. This paper demonstrates that deep learning (CNN) guided adaptive key generation, coupled with robust permutation-diffusion architecture and non-chaotic (PRNG) masking, can deliver outstanding theoretical and empirical security in image encryption even without traditional chaotic maps. Our CNN guided enhancement dynamically generates permutation vectors based on image features, introducing data adaptivity, larger key space, and enhanced security. This approach opens avenues for secure multimedia protection leveraging the frontier of AI driven cryptography and also reduces the time complexity. The proposed system achieves near ideal entropy, superior NPCR/UACI ratios, and high resilience to differential and statistical attacks while preserving computing efficiency suitable for real time systems. Experimental validation exhibits improvement over previous works in many instances.