Background: Offline signature verification is challenging due to high variabilityand limited data. While MobileNetV3-Small enables on-device use, models pretrained on ImageNet lack handwriting-specific features. This study explores a threestage transfer learning approach: first adapting MobileNetV3-Small to Devanagaricharacter recognition, then fine-tuning on BHSig260-Hindi signatures, followed byhyperparameter tuning.Objectives: The primary goal is to assess whether handwriting-focused transfer learning improves signature verification performance compared to the standardImageNet-only baseline. Specifically, the study aims to (1) adapt MobileNetV3-Smallto Devnagari characters, (2) fine-tune the adapted backbone for genuine-vs-forgedsignature classification, (3) optimize the classification head and learning rate throughhyperparameter tuning, and (4) evaluate gains in accuracy, precision, recall, F1-score,and AUC. Methods: In Stage 1, MobileNetV3-Small (include_top=false) was pretrained ona Devanagari handwritten character dataset (49 classes, 8,273 images), using standard augmentations and freezing 80% of layers. In Stage 2, the model was fine-tunedon the BHSig260-Hindi dataset (genuine vs. forged) with a binary classifier head. InStage 3, hyperparameter tuning was performed using Keras Tuner (RandomSearch)to optimize dense units, dropout, and learning rate.Results: Domain-aware transfer learning led to consistent improvements over theImageNet-only baseline. After tuning, the final model achieved an average accuracy of 94.52%, F1-score of 94.99%, and AUC of 98.56% on the test set, withresults stable across multiple runs. However, since the baseline model was nothyperparameter-tuned, the comparison is not entirely fair; the goal was to explorethe potential benefits of handwriting-based pretraining and tuning when applied toMobileNet for offline signature verification.Conclusions: Fine-tuning MobileNetV3-Small on a structurally similar handwriting dataset improves its ability to distinguish genuine from forged signatures. Theresults show that lightweight CNNs benefit from handwriting-aware transfer learning and targeted hyperparameter tuning, making them well-suited for accurate andefficient on-device signature verification.Keywords: Handwritten Signature Verification; Transfer Learning; Hyperparameter Tuning; MobileNetV3-Small; Devnagari Characters; BHSig260-Hindi.