Answers are grounded in this paper and cite their sources.
Smart-grid intrusion detection faces three coupled challenges: generalization to unseen events, unequal evidence quality across modalities, and false alarms caused by sensor perturbations. Random sample-level splitting can produce optimistic estimates when neighboring samples from the same event occur in both the training and test sets. This paper presents the Cross-Modal Evidence Conflict Fusion model (CMECF), which combines modal credibility, pairwise evidence conflict, conflict-conditioned message passing, and encoder-level physics-consistency constraints over four synchronized evidence streams: PMU measurements, control-panel logs, protective-relay logs, and Snort alerts. On the MSU/ORNL binary dataset, Random Forest recorded macro-F1 of 0.896 under a random split and 0.564 under an event-level split. In the seed-42 event-level evaluation, CMECF recorded macro-F1 of 0.595, weighted-F1 of 0.623, and FPRA of 0.604; the corresponding Random Forest macro-F1 and FPRA were 0.564 and 0.722. Under the PMU Gaussian perturbation at σ =0.3 , macro-F1 was 0.523 for CMECF and 0.378 for Random Forest. CMECF has approximately 105\,K parameters and a model-forward latency of 1.5 ms per sample on the stated hardware. All comparative results come from a single seed-42 run and are reported descriptively. Within this evidence, CMECF records the highest observed macro-F1 and weighted-F1, the lowest measured false-positive rate, stronger performance retention under the evaluated PMU perturbation, and a compact millisecond-scale four-modality inference profile.
No takes yet. Share an insight, caveat, or question.
Yin et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: