In recent years, many studies have focused on the robustness of federated learning (FL) systems against Byzantine attacks. Prior work has shown that integrating stochastic optimization algorithms with momentum can effectively enhance the robustness, particularly in non-convex scenarios. However, most studies assume a fixed batch size across all workers, which may not be suitable for practical deployments. In this study, we investigate the impact of different batch sizes across different clients on Byzantine robustness in FL. Specifically, we propose a weighted aggregation framework (WAggF), which assigns different weights to gradients received from different clients. WAggF can be easily embedded into various aggregation rules to boost Byzantine resilience. We further establish theoretical convergence guarantees for two momentum-based stochastic optimization algorithms—Byz-WNSGDM and Byz-EF21-WSGDM—under WAggF. To reduce communication overhead and enable synchronous training, we incorporate a contractive compressor (which encompasses Top-K as a special case) into the stochastic optimization algorithm. Extensive experiments demonstrate that our proposed algorithms consistently improve robustness against a range of Byzantine attacks.
Artificial Intelligence (AI)-enabled systems are now increasingly deployed across safety-critical, mission-critical, and socio-technical domains. Existing engineering disciplines offer mature approaches to human and software assurance; however, AI-enabled systems can exhibit characteristics such as probabilistic behaviour, data dependency, limited explainability, and adaptation that challenge traditional assurance methods. Our research examines assurance precepts across three domains that arguably dominated assurance reforms in capability development during different periods: (1) human assurance, circa 1980 to 2000, (2) software assurance, circa 2000 to 2020, and (3) assurance of AI-enabled systems, emerging since 2020. While software assurance is largely concerned with verifying deterministic behaviour against specified requirements and human assurance focuses on managing variability in socio-technical systems, AI assurance must address systems whose behaviour is learned, probabilistic, and inherently dependent on data. We contribute a set of 20 assurance precepts for safety-critical AI-enabled capabilities. To extend current assurance frameworks, we structured these precepts across four quadrants of assurance activities, responsibilities, and critical questions, and organized them within a novel Dual Assurance Spiral for AI-enabled Capabilities (DAS4AIC). We propose the DAS4AIC to operationalise the 20 AI assurance precepts throughout capability development and sustainment. Further, we demonstrate how classical safety precepts map from human-dominant assurance through software-dominant assurance to the assurance of AI-enabled systems with conceptual validation. Our preliminary evaluation is that some human assurance precepts map more directly to AI-enabled systems than through the intermediate step of software assurance, suggesting that established human assurance concepts may have continuing relevance for AI-enabled capabilities. To our knowledge, this is the first study to align emerging AI-assurance precepts explicitly with the earlier and still overlapping, human-dominant and software-dominant assurance paradigms.
Residue-soil mixing and reduced spectral sensitivity complicate summer maize leaf area index (LAI) estimation from unmanned aerial vehicle multispectral imagery. We compared full-plot and canopy domains, paired differences, neighborhood statistics, and approximate days after sowing (DAS), calculated from a common reference date. Five input schemes (B0–B4) and four algorithms were evaluated using 345 samples collected on five dates in one growing season. Spatial validation comprised ten repetitions of five-fold grouped cross-validation, with preprocessing and feature selection restricted to training data. For gradient boosting decision trees, R2 increased from 0.515 for B0 to 0.686 for B3 and 0.846 for B4, whereas RMSE decreased from 0.834 to 0.671 and 0.470. Low-LAI Bias decreased from 0.580 to 0.343 and 0.160. Fixed-capacity controls supported conditional improvements after adding paired differences and neighborhood statistics. Approximate DAS produced the largest sequential RMSE reduction, mainly through between-date adjustment. Time-baseline and training-fold residual analyses distinguish calendar-time adjustment from within-date predictive information. High-LAI underestimation and unseen-date errors delimit application beyond the sampled dates and production-field conditions. Within this defined scope, the framework provides a reproducible comparison of image-domain representations under heterogeneous production conditions.
Although machine learning-based forecasting models often function as opaque black boxes, explainability methods provide the essential tools required to foster trust in deep learning systems, optimize computational resources, and enhance model accuracy. Motivated by this, we propose a model-agnostic, local explainability method specifically designed for forecasting tasks. Our approach employs a hierarchical structure that is generated iteratively. From the outset, a small number of “interpretable components” that decompose the input are considered. In the inductive step, the methodology enables analysts to examine the data in greater detail, allowing for a deeper exploration of specific subcomponents whenever a higher resolution is required. This framework inherently provides control over computational costs by tuning the branching factor. The resulting explainability tree shares certain, but not all, properties of the explainability values provided by the SHAP-LIME framework; in this regard, we present some theoretical results. Furthermore, numerical experiments provided evidence that explanation coherence is fundamentally maintained between both methods. We explored the utility of this methodology on a real-world short-term load forecasting LSTM encoder–decoder model.
Warfarin dosing is complicated by wide inter-patient variability driven by genetic, demographic, and pharmacological factors. Although machine learning approaches to dose prediction have been widely studied, published evaluations are frequently affected by analysis-domain data leakage, small held-out sets, and an absence of calibration and fairness assessment across ancestry subgroups. We describe WarfaRisk, a nine-phase open-source machine learning pipeline for warfarin dose prediction evaluated on two public International Warfarin Pharmacogenetics Consortium (IWPC) cohorts (n=6037 and n=1780, the latter a non-independent secondary consistency check) using fixed patient-level splits and a single documented random seed. The pipeline incorporates leakage-free preprocessing, nested clinical and pharmacogenomic feature ablation across nine model architectures, leave-one-ancestry-out portability assessment with a cross-model robustness check, conformal prediction calibration reported at subgroup resolution, and individual-level SHapley Additive exPlanations (SHAP) feature attribution. The nine-architecture ablation results reported here are from the primary cohort, IWPC-6256; the complete two-cohort comparison is reported in a companion evaluation manuscript. A prototype web-based clinical decision support interface integrates dose recommendation, uncertainty quantification, and SHAP explainability with a Fast Healthcare Interoperability Resources (FHIR) R4 endpoint. In preliminary evaluation on a single fixed held-out split, Model Agnostic Prediction Interval Estimator (MAPIE) cross-conformal (5-fold CV+) prediction achieves coverage close to its nominal target when measured across each cohort as a whole. When this coverage is instead stratified by ancestry, a different pattern emerges: coverage for African-ancestry patients falls well short, a gap the aggregate number hides, plausibly explained by alleles missing from the standard IWPC genotype panel and supported by independent pharmacogenetic literature, though not directly demonstrated in this dataset. These findings indicate that subgroup-level calibration reporting is necessary, not optional, since its absence can let a model appear well-calibrated in aggregate while failing a specific population in practice.
