GreenSense: controlled predictive, computational, and carbon-aware evaluation of efficient neural models for fraud detection
Year: 2026
Type: article
Abstract: Green artificial intelligence requires model quality to be evaluated together with computational cost, model size, and estimated environmental impact. This study presents GreenSense as a controlled benchmarking framework for energy-aware fraud detection. Seven configurations are compared on two imbalanced transaction datasets: Compact Nonlinear Dense (CND), full-precision and dynamically quantized tabular CNNs and MLPs, an LTC-style state-refinement network, and a sparse NCP-CfC-style network. All training, quantization, and inference measurements were repeated on a single dedicated workstation (Intel Core i5-13500, Windows 11) under one fixed software stack, thread and CPU-affinity setting, and measurement protocol. Each model–dataset configuration is evaluated over five independently seeded training runs, and each inference measurement is repeated five times in fixed-work windows of at least 30 s. CPU-package energy is measured with CodeCarbon and cross-checked against a second software readout of the same hardware energy counter; CO $$_2$$ is derived from a cited national grid-emission factor. Predictive evaluation emphasizes precision–recall area under the curve (PR-AUC), $$F_1$$ , recall, precision, and ROC-AUC. On FraudTrain, the LTC-style model achieved the highest mean PR-AUC (0.8015, 95% CI 0.7947–0.8082), followed by MLP-FP32 (0.7679) and NCP-CfC-style (0.7654). On the ULB Credit Card benchmark, MLP-FP32 (0.7554) and CND (0.7528) had the highest mean PR-AUC with overlapping intervals, while CNN-FP32 had the highest mean $$F_1$$ (0.1889). Training accounted for more than 99.9% of lifecycle-equivalent energy for every configuration. Dynamic INT8 quantization reduced MLP storage size by 1.94 $$\times $$ (FraudTrain) and 2.09 $$\times $$ (ULB) but increased inference energy per transaction in all four FP32–INT8 pairs (by 43–47% for the MLP and 4–6% for the CNN; all paired intervals excluded zero). Its predictive effect was dataset-dependent: FraudTrain MLP PR-AUC decreased by 0.301 (95% CI $$-0.358$$ to $$-0.247$$ ), whereas the ULB change was only $$-0.003$$ . The PR-AUC/lifecycle-energy Pareto sets were CND, MLP-FP32, and LTC-style on FraudTrain and CND and MLP-FP32 on ULB. GreenSense provides a reproducible, task-specific basis for comparing predictive, computational, and estimated environmental operating points; all code, configurations, raw energy logs, and manifests are archived with a DOI.
Source: Scientific Reports
Authors: K. Sathya(Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Visualization, Writing – original draft, Writing – review & editing), M. Pandiyan(Conceptualization, Data curation, Formal analysis, Methodology, Writing – original draft, Writing – review & editing), S. Tamilselvi, Nitheesh Murugan Kaliyamurthy(Investigation, Software, Visualization, Writing – original draft, Writing – review & editing), Swapnesh Taterh, Saurabh Agarwal(Supervision), Shashi Kant(Supervision)
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Climate action
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7234259455
