Paper 2025/688

Uncertainty Estimation in Neural Network-enabled Side-channel Analysis and Links to Explainability

Abstract

Side-channel analysis (SCA) has emerged as a critical field in securing hardware implementations against potential vulnerabilities. With the advent of artificial intelligence(AI), neural network-based approaches have proven to be among the most useful techniques for profiled SCA. Despite the success of NN-assisted SCA, a critical challenge remains, namely understanding predictive uncertainty. NNs are often uncertain in their predictions, leading to incorrect key guesses with high probabilities, corresponding to a higher rank associated with the correct key. This uncertainty stems from multiple factors, including measurement errors, randomness in physical quantities, and variability in NN training. Understanding whether this uncertainty arises from inherent data characteristics or can be mitigated through better training is crucial. Additionally, if data uncertainty dominates, identifying specific trace features responsible for misclassification becomes essential. We propose a novel approach to estimating uncertainty in NN-based SCA by leveraging Renyi entropy, which offers a generalized framework for capturing various forms of uncertainty. This metric allows us to quantify uncertainty in NN predictions and explain its impact on key recovery. We decompose uncertainty into epistemic (model-related) and aleatoric (data-related) components. Given the challenge of estimating probability distributions in high-dimensional spaces, we use matrix-based Renyi α-entropy and α-divergence to better approximate leakage distributions, addressing the limitations of KL divergence in SCA. We also explore the sources of uncertainty, e.g., resynchronization, randomized keys, as well as hyperparameters related to NN training. To identify which time instances (features in traces) contribute most to uncertainty, we also integrate SHAP explanations with our framework, overcoming the limitations of conventional sensitivity analysis. Lastly, we show that predictive uncertainty strongly correlates with standard SCA metrics like rank, offering a complementary measure for evaluating attack complexity. Our theoretical findings are backed by extensive experiments on available datasets and NN models.