Ransomware, particularly crypto-ransomware, remains a significant cybersecurity challenge, encrypting victim data and demanding a ransom, often leaving the data irretrievable even if payment is made. This study proposes an early detection approach to mitigate such threats by identifying ransomware activity before the encryption process begins. The approach employs a two-tiered approach: a signature-based method using hashing techniques to match known threats and a dynamic behavior-based analysis leveraging Cuckoo Sandbox and machine learning algorithms. A critical feature is the integration of the most effective Application Programming Interface call monitoring, which analyzes system-level interactions such as file encryption, key generation, and registry modifications. This enables the detection of both known and zero-day ransomware variants, overcoming limitations of traditional methods. The proposed technique was evaluated using classifiers such as Random Forest, Support Vector Machine, and K-Nearest Neighbors, achieving a detection accuracy of 98% based on 26 key ransomware attributes with an 80:20 training-to-testing ratio and 10-fold cross-validation. By combining minimal feature sets with robust behavioral analysis, the proposed method outperforms existing solutions and addresses current challenges in ransomware detection, thereby enhancing cybersecurity resilience.

Ransomware is malware that encrypts data without permission, demanding payment for access. Detecting ransomware on Android platforms is challenging due to evolving malicious techniques and diverse application behaviors. Traditional methods, such as static and dynamic analysis, suffer from polymorphism, code obfuscation, and high resource demands. This paper introduces a multi-stage approach to enhance behavioral analysis for Android ransomware detection, focusing on a reduced set of distinguishing features. The approach includes ransomware app collection, behavioral profile generation, dataset creation, feature identification, reduction, and classification. Experiments were conducted on ∼3300 Android-based ransomware samples, despite the challenges posed by their evolving nature and complexity. The feature reduction strategy successfully reduced features by 80%, with only a marginal loss of detection accuracy (0.59%). Different machine learning algorithms are employed for classification and achieve 96.71% detection accuracy. Additionally, 10-fold cross-validation demonstrated robustness, yielding an AUC-ROC of 99.3%. Importantly, latency and memory evaluations revealed that models using the reduced feature set achieved up to a 99% reduction in inference time and significant memory savings across classifiers. The proposed approach outperforms existing techniques by achieving high detection accuracy with a minimal feature set, also suitable for deployment in resource-constrained environments. Future work may extend datasets and include iOS-based ransomware applications.