Crypto systems deployed in smart cards and embedded devices are vulnerable to side-channel attacks, in which adversaries exploit power consumption to recover secret keys. Correlation power analysis (CPA) and machine-learning-based attacks can break AES-128 implementations with relatively few traces. This work investigates Neural Signal Obfuscation (NSO) as a practical countermeasure for an AES-128 smart card-style implementation. We first acquire and characterize power traces from the target device and quantify baseline leakage using traces-to-disclosure (TTD), guessing entropy (GE), signal-to-noise ratio (SNR), and the success rate of a neural network attacker. We then design and train a lightweight one-dimensional convolutional obfuscator that transforms power traces to suppress exploitable leakage while preserving correct cryptographic functionality. Re-evaluating obfuscated traces under the same CPA and machine-learning attacks shows that NSO increases TTD, reduces GE and SNR, and lowers the attacker's accuracy, forcing an adversary to collect significantly more traces to recover the key than in the unsecured baseline implementation.