method: Continual-Positive Congruent Training2025-03-28
Authors: Daniele Ghiani
Affiliation: University of Cagliari
Description: Continual-Positive Congruent Training (C-PCT) is an adaptation of the well-known PCT strategy to a Continual Learning (CL) scenario.
In CL, data arrives incrementally, following a temporal order. Each time a new batch of data becomes available, the model is updated using only the most recent data.
Following this scenario the training data has been divided into 12 quarters.
The PCT regularization loss term encourages the new model under training, which is updated solely with the latest data, to mimic the previous model's behavior when it made correct predictions.
method: Baseline - SecSVM2025-03-10
Authors: .
Affiliation: .
Description: Reproduction of the detector from Demontis et al. “Yes, machine learning can be more secure! a case study on android malware detection.” IEEE TDSC 2017.
method: SVM-CB (b=0.2, n=100)2025-03-31
Authors: Angioni, Daniele; Demetrio, Luca; Pintor, Maura; Biggio, Battista
Affiliation: University of Cagliari
Description: Support Vector Machine with Custom Bounds (SVM-CB). It first trains a baseline Linear SVM to compute the T-stability measure for each feature. It then uses this information to train another Linear SVM enforcing a hard constraint on the weights of the features with the most negative T-stability values.
Angioni et al. "Robust Machine Learning for Malware Detection over Time." ITASEC 2022: 169-180
| Detection Rate | False Positive Rate | ||||||
|---|---|---|---|---|---|---|---|
| Date | Method | 100 manipulated features | Clean data | False Positive Rate | |||
| 2025-03-28 | Continual-Positive Congruent Training | 35.84% | 59.20% | 0.40% | |||
| 2025-03-10 | Baseline - SecSVM | 11.68% | 75.36% | 0.44% | |||
| 2025-03-31 | SVM-CB (b=0.2, n=100) | 4.88% | 75.36% | 0.82% | |||
| 2025-02-17 | Baseline - DREBIN | 3.76% | 77.28% | 0.36% | |||