Accurate and robust min-entropy evaluation is essential for assessing the security of random number generators, yet existing learning-based estimators still face two major limitations. First, most methods fix the prediction target at 8 bits, although the target bit width directly determines the granularity of the induced prediction space and can substantially affect estimation accuracy. Second, many methods are evaluated only on either independent and identically distributed sources or correlated non-independent sources, which limits the evidence of robustness across heterogeneous statistical regimes and leaves their behavior on real-world black-box sources insufficiently understood. To address these issues, this paper first presents a theoretical analysis of adaptive target-bit selection and shows that a fixed target granularity can lead to a non-negligible loss of accuracy, whereas profiling target bit widths from 1 to 16 provides a more principled basis for identifying a suitable operating point. On this basis, we propose RobeTimeVis, a cross-modal semantic–visual framework for min-entropy evaluation. RobeTimeVis combines a RoBERTa-based semantic encoder for modeling sequence-level dependence, a lightweight Tiny-ViT visual encoder for extracting structural information from multi-channel time-series images constructed by GAF, MTF, and RP, and a cross-modal attention module for representation fusion. Experiments on 18 simulated entropy sources with known theoretical min-entropy show that RobeTimeVis improves both accuracy and robustness across heterogeneous statistical regimes. Aggregated over all datasets and target bit widths, the mean absolute error decreases from 3.453 bits for NIST SP 800-90B to 0.722 bits. On generalized binary autoregressive processes and high-degree m-sequences, the reductions are from 6.379 to 0.017 bits and from 6.785 to 0.566 bits, respectively. We further evaluate the framework on real-world Random.org data under a black-box setting with unknown min-entropy, using a practical conservative black-box evaluation criterion. The results show that the same adaptive profiling framework remains applicable beyond controlled benchmarks and supports deployment-oriented assessment when ground truth is unavailable.
Feng et al. (Wed,) studied this question.