[Paper Link (will be provided soon)]
Jimin Park, Haeseung Jeon, Ahyun Ji, Aritran Piplai, Mohammad Saidur Rahman, Se Eun Oh†
†Corresponding author.
Note
This is the authors' TabCL attack model code proposed in TabCL: Continual Malware Classification with Tabular-Aware Generation work, presented in the PAKDD'26.
The pipeline of TabCL training
A CTGAN-based generative model and a classifier were jointly trained in an incremental learning setting
- Initial number of training classes: 50, with 5 new classes added per step, reaching 100 classes in total
- Total number of training tasks: 11
Training Schedule
- First task: 25 epochs
- Each subsequent task: 5 epochs
CTGAN Architecture and Hyperparameters
- Latent vector dimension: 128
- Generator and Discriminator structure: MLP with hidden layers [256, 256]
- Learning rate (Generator & Discriminator): 2 × 10⁻⁴
- Weight decay: 1 × 10⁻⁶
- Mini-batch size: 256
- Update ratio: 1 discriminator update per generator update
Classifier Training Setup
- Optimizer: SGD
- Learning rate: 1 × 10⁻³
- Momentum: 0.9
- Weight decay: 1 × 10⁻⁶
- Mini-batch size: 256