Institutional Repository of Vilnius University
Tęstinio mokymosi taikymas kenkimo programų klasifikacijai naudojant giliuosius neuroninius tinklus /
Abstract
dc:descriptionWith the growing volume of new malware, deep neural networks are increasingly used for malware detection and classification. However, updating these models in a continual learning setting can lead to catastrophic forgetting. The aim of this work is to compare the effect of elastic weight consolidation, experience replay, generative replay, and a combination of elastic weight consolidation and generative replay on the stability-plasticity balance of a deep neural network applied to malware classification. For the experiments, a binary file dataset was constructed with nine malware families and one benign class. The compared methods were EWC, utilising true FIM; experience replay with a herding algorithm; DGR with GAN-generated pseudo-samples; a combined EWC-DGR approach. The results show that experience replay achieved the best stability-plasticity balance. EWC preserved first-task knowledge more effectively, but stronger regularization limited the model’s ability to learn new tasks. Combining EWC with DGR reduced this effect and often preserved second-task accuracy better than EWC alone. The DGR results also indicate that GAN-generated pseudo-samples were less effective than real samples, and that quantitative GAN evaluation metrics do not necessarily reflect the usefulness of pseudo-samples for continual learning.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Cicėnaitė, Vesta,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308336465&prefLang=en_US