Unifying temporal continuity and structural decomposition for stealthy FDIA detection in smart grids via implicit neural representations

False data injection attacks (FDIAs) pose significant threats to power system security by stealthily manipulating measurements to bypass traditional residual-based detectors. While recent time-series detection methods improve robustness by modeling temporal continuity, they often rely on labeled attack data or static architectures with limited adaptability. This paper proposes a fully unsupervised FDIA detection framework based on implicit neural representations (INRs), which model multivariate power system measurements as continuous functions of time. A transformer-based hypernetwork dynamically generates INR parameters, enabling the model to adapt to varying operating conditions without retraining. To unify temporal continuity and structural decomposition, the proposed framework couples continuous-time INR modeling with a hierarchical trend-seasonal-residual representation and a measurement-type-aware residual structure. Anomalies are identified via a weighted reconstruction error using type-normalized anomaly scores. Extensive experiments on IEEE 14-bus and 118-bus test systems demonstrate that the proposed method achieves near-perfect detection accuracy against both optimization-based and controllable FDIAs, significantly outperforming existing supervised, semi-supervised, and unsupervised baselines. Moreover, the INR model exhibits strong sensitivity to subtle, temporally-localized perturbations and remains robust under different attack amplitudes. To the best of our knowledge, this is the first FDIA detection framework leveraging INR-based continuous-time modeling with transformer-driven parameterization, offering a scalable, interpretable, and topology-agnostic solution for securing modern power grids. • Introduces the first INR-based FDIA detection framework for power system measurements. • Embeds a learnable trend-seasonal-residual decomposition within the INR to provide multi-scale inductive bias and enhance interpretability. • Demonstrates near-perfect detection accuracy on IEEE benchmark systems under both optimization-based and controllable FDIAs in a fully unsupervised and topology-agnostic setting.