Artificial Intelligence Adoption and Corporate Financial Performance: Evidence from Indian Firms

The detection involves correctly categorizing the smishing message and delivering the warning clearly to facilitate mobile cybersecurity decision making. In this study, an explainable machine-learning pipeline is evaluated on a public dataset of 5,971 legitimate, spam and smishing messages. After removing conflicting labels and repeated messages, the number of records to be used in the training, validation and test sets was 5,797. Multiple classifiers were used to assess word and character TF-IDF representations with structural attributes. The logistic regression model trained with text-only data, selected by training cross-validation, had an accuracy of 97.36% and an F1 score of 0.9156 (macro average). When a warning threshold of 0.45 was used, precision and recall for smishing were 93.42% and 89.87% respectively. The mean maximum predicted-class probability reduction using token-deletion analysis for 30 messages was 0.0942, and four changes in classification occurred. The experimental environment had a median prediction latency of 1.22 milliseconds. The conclusions are relevant to efficient and inspectable SMS screening, and for operational effectiveness and the usefulness of explanations, external validation, probability calibration, and SMS testing with smartphones are still required.