Counterfeit currency remains a persistent threat to financial systems, and conventional verification based on manual inspection or fixed-rule hardware often fails under real-world conditions such as poor illumination, worn notes, and increasingly sophisticated forgeries. Most existing deep learning solutions rely solely on visible RGB imagery and therefore miss material- and ink-level cues that are observable only outside the visible spectrum. This paper proposes a dual-branch deep learning framework that fuses spatial features from RGB images with spectral features from ultraviolet (UV) and infrared (IR) imaging through an attention-guided intermediate fusion mechanism. The complete architecture, fusion formulation, training objective, and a four-way comparison of candidate fusion strategies (early, late, intermediate, and attention-based) are presented, together with an evaluation protocol benchmarking the approach against RGB-only, spectral-only, and classical handcrafted-feature baselines. Two additional metrics, Specificity and Matthews Correlation Coefficient (MCC), are introduced alongside standard accuracy-oriented and security-oriented indicators (FAR/FRR) to give a more complete picture of classification balance. Reported trends indicate that spectral-spatial fusion improves robustness under low-light, blur, and noise conditions relative to single-modality models. Limitations related to dataset scale, sensor dependency, and cross-currency generalization are discussed along with future directions in explainable AI, few-shot learning, and federated training.

Counterfeit currency detection; spectral-spatial fusion; deep learning; attention mechanism; UV/IR imaging; convolutional neural network.