Deep learning–based power analysis attacks pose increasing challenges to the security of cryptographic systems in embedded and IoT devices. This work presents a countermeasure for AES-128 that combines compact \(T-box\) lookup tables with a convolutional neural network (CNN)-based key scheduling mechanism to enhance resistance against such attacks. The proposed approach introduces two core components: (1) 8 × 8-bit merged MixColumns-SubBytes tables that reduce data-dependent power leakage, and (2) a non-linear CNN-based key scheduler designed to disrupt recognizable trace patterns used in side-channel analysis. Experimental results show that this design increases the number of traces required for successful key recovery to at least 3,900, compared to 1,100 traces in the unprotected implementation, while introducing only a 2 × runtime overhead. The method also achieves a 72% improvement in resistance to correlation power analysis attacks over standard masking techniques. These contributions offer a practical solution for strengthening cryptographic implementations in resource-constrained environments.
