Abstract
Deep learning detectors of false data injection (FDI) attacks in smart grids are often reported with near-perfect accuracy. This study evaluates convolutional (CNN), long short-term memory (LSTM), and CNN–LSTM detectors on generic, non-stealthy bias, scale, and spike perturbations injected into 60,000 operating points of decentral smart grid control (DSGC) stability data, grouped into 20-step pseudo-temporal windows. Windows with at least one attacked point were labeled attacked, so a rule that always predicts an attack reaches 97.83% accuracy. In one run per model, the CNN reached 98.48% accuracy but a false-positive rate (FPR) of 34.4% and a Matthews correlation coefficient (MCC) of 0.64, whereas the LSTM (FPR 4.4%, MCC 0.81) and the CNN–LSTM (FPR 7.9%, MCC 0.81) did better. On a second, fully specified benchmark, a one-line power-balance check raised no false alarms (MCC 0.35). Unweighted window-level detectors stayed near the always-attack rule until most attacked values left their ranges, where a range check did better; a row-level network learned the balance at the smallest magnitude. Small balance-keeping shifts escaped all detectors. SHAP, LIME, and integrated gradients located large attacks (AUROC up to 0.97) worse than a model-free deviation (0.98). Results concern evaluation practice on a static benchmark, not real grids.