Comparative Evaluation of CPA and Machine Learning-based power analysis attacks on AES, PRESENT, and ASCON

Advanced Encryption Standard (AES), PRESENT, and ASCON are widely deployed encryption algorithms in various applications ranging from servers to resource-constrained devices. This paper applies classical Correlation Power Analysis (CPA) and Machine Learning (ML) based power analysis attacks to all three ciphers. Power traces are captured with a ChipWhisperer-Lite from an 8-bit ATXMEGA microcontroller running each cipher. Subsequently, to emulate the practical side-channel countermeasures and assess the robustness of these methods, trace misalignment (jitter) and amplitude noise at controlled levels are injected into these clean captured traces. CPA and four ML models are employed and compared in three different phases: First using clean traces; second, using traces with either noise or jitter; and the last phase uses traces where both noise and jitter have been combined. A comprehensive performance evaluation is conducted by measuring key recovery rate, accuracy, F1 score, and runtime. On the clean traces, every method recovers the full key for AES and PRESENT ciphers. However, ASCON resists standard CPA, so it is attacked with second-order (SO) CPA, which recovers up to 50% of the key bytes. Under distortion, no single method holds up in every case: CPA and Random Forest tolerate noise but fail under jitter, while SVM tolerates jitter but collapses under noise. Overall, the ML power analysis attacks outperform CPA under most noise and misalignment conditions, whereas the phase three evaluation confirms that combining noise with jitter defeats nearly all of the evaluated attacks.