Cybercrime Forecasting Using Convolutional Neural Network-Long Short-Term Memory Networks

In today's digital world, cybercrime has evolved into a global challenge, with threats such as ransomware, data breaches, identity theft, and financial fraud affecting organizational networks and individual data.Conventional digital forensics methods primarily adopt a reactive approach in which evidence acquisition begins after a network breach has occurred.This inherent latency may increase the risk of losing critical information and can limit the effectiveness of post-incident analysis.To address this limitation, this study proposes a proactive forecasting approach for predicting the subsequent state of network activity from preceding observations.A hybrid learning architecture integrating Convolutional Neural Networks and Long Short-Term Memory networks is employed to model relationships within network-flow features and support one-step-ahead prediction.The framework is evaluated on the UNSW-NB15 benchmark and independently validated on the CIC-IDS2017 benchmark to assess its applicability beyond a single dataset.On UNSW-NB15, the model achieved an accuracy of 96.07%and a recall of 96.18%.On CIC-IDS2017, the model achieved 96.56% accuracy, 82.78% precision, 78.37% recall, an F1-score of 80.52%, and an area under the receiver operating characteristic curve of 0.9051.These findings provide additional evidence that the proposed Convolutional Neural Networks and Long Short-Term Memory forecasting framework can be applied across different network-security benchmarks, while also indicating that predictive performance varies according to dataset characteristics.