Paper 2024/071

Too Hot To Be True: Temperature Calibration for Higher Confidence in NN-assisted Side-channel Analysis

Abstract

The past years have witnessed a considerable increase in research efforts on neural network-assisted profiled side-channel analysis (SCA). At the same time, studies have identified challenges, including closing the gap between machine learning (ML) classification metrics and side-channel attack evaluation. In fact, in NN-assisted SCA, the NN’s output distribution forms the basis for successful key recovery. In this respect, prior work has studied many aspects of integrating NNs into SCA, including model selection, training, and hyperparameter tuning. Nevertheless, a well-known fact has been largely overlooked in the SCA-related literature, namely NNs’ tendency to become over-confident, that is, assigning overly high probability to the correct class (the secret key in the sense of SCA). Temperature scaling is a powerful remedy for this behavior. From the perspective of deep learning, temperature scaling does not affect NN accuracy; however, its impact on secret-key recovery metrics, mainly guessing entropy, is worth investigating. This paper reintroduces temperature scaling into SCA and demonstrates that key recovery can become more effective. Importantly, temperature scaling can be integrated into SCA without re-tuning the network. In doing so, temperature can be treated as a metric to assess NN performance before launching the attack. In this regard, we study the impact of hyperparameter tuning, network variance, and capacity, and derive recommendations to prevent miscalibration and overconfidence.