Answers are grounded in this paper and cite their sources.
Many attacks come from trusted users in the organization. As the nature of user activities in cyber systems is irregular and unpredictable, user-related threats are always an issue in cyber networks. As a result, many approaches have been proposed for the detection of outliers in user behavior. To overcome the above problems, we consider TL-CsLeNet (Transfer Learning-based Customized LeNet) framework for anomaly detection is evaluated and refined comprehensively. The assessed framework consists of three main phases, namely, pre-processing using Slice Adaptive Sigmoid Z-score (SASZ) normalization, behavioral feature generation and deep architectural classification. To guarantee methodological rigor, we study an imbalanced public UEBA dataset (528,690 events; 14.25% anomaly rate) in a systematic 17-configuration component study. The most important result is that the precision-recall area (PR-AUC) is 0.22 lower if all engineered features are removed, and 0.18 lower if the HOS features alone are omitted, a factor of an order higher than any of the architectural components. The results indicate that a refined configuration, employing standard Information Gain, max pooling and no transfer learning (and selected only by using validation data) yielded an accuracy of 91.76% and a precision of 0.7421 (averaged over four training splits and three random seeds). It is statistically equivalent to the best comparison architectures (LeNet, LinkNet and DCNN) and much superior to the other five architectures; it is also as stable as the most stable baselines. This component study also demonstrates that the Modified Deng Entropy–Information Gain (MDE-IG) filter, pyramid pooling, mixed pooling, and SiLU activation don’t have any measurable impact, and transfer learning negatively impacts performance. From these results, we can conclude that feature construction is one of the most important things to do when it comes to quality data on tabular cybersecurity logs, and the results verified this statement. Clinical trial number : Not applicable.
No takes yet. Share an insight, caveat, or question.
Naaz et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: