Genshield AI: Intelligent Ransomware Detection and Prevention Using Generative AI

DOI

https://www.doi.org/10.59256/indjcst.20260503082

Published In

Volume 5, Issue 3 (2026)

Published Online

September-December 2026

Pages

658-676

Article Type

Original Article

Check for updates

When information leakage or testing is weak, the accuracy of ransomware detection systems will be very high. This study explores the reimagined GENSHIELD AI, a multi-layered security offering that utilizes autoencoder anomaly detection with supervised machine learning, the Vision Transformer (ViT) family class attribution approach and a moderated connoisseur Generative AI advisory. The datasets used for testing the system were two public datasets: 62,485 records from MLRD and 58,596 records from CIC-MalMem-2022. After cleaning, the MLRD was found to have 32,240 unique feature vectors and out-of-scope testing yielded 39,089 detection records and 19,507 other-malware records from CIC-MalMem-2022. With benign fitted log-standard scaling and 99th percentile threshold, the autoencoder achieved a perfect recall, F1 score of 0.986 and an accuracy of 0.991 in CIC-MalMem-2022 while having a false-positive rate of 0.009. On MLRD, however, only 39.5% was recalled while precision was 0.952. Duplicate data contributed to the strong performance and low observed false-negative rate of the supervised models. With a macro-F1 score of 0.335 for the ViT, family attribution was less certain. Additionally, the memory model detected 100% of the 19,507 malware and trojan entries it came across. Additionally, the memory model flagged 100% of spyware/trojan data for 19,507, indicating a limited detection of ransomware. By eliminating active content and adding a tiny 0.3ms processing time, the guarded Generative AI layer became effective. GENSHIELD AI is more suitably considered as a malware triage system and not as a one-shot ransomware solution for the time being, pending a more extensive deployment and validation.