Research map / Continual learning / Online and replay-based
MalCL: Leveraging GAN-Based Generative Replay to Combat Catastrophic Forgetting in Malware Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MalCL, a GAN-based generative replay continual learning system for malware family classification that uses feature matching loss and logit-based replay sample selection.
Problem
Catastrophic forgetting in continual learning for malware family classification, where classifiers lose accuracy on previously seen malware families as new families arrive in a class-incremental setting.
Approach
Combines a GAN generator trained with feature matching loss to synthesize replay samples with selection schemes that pick synthetic samples closest to per-class mean logits of the classifier's hidden layer, tailored to malware data rather than image-domain generative replay.
Key results
- MalCL with FML and L1 to per-class mean logits reaches 54.5% mean accuracy on EMBER over 11 tasks, versus 26.8% for GR and 27.0% for BI-R and 27.5% for the None baseline.
- With larger classes assigned to the first task, MalCL reaches a mean accuracy of 74% on EMBER, against a Joint baseline of 88.7%.
- FML slightly outperforms BCE across tasks for both selection schemes, and L1 to per-batch mean logits performs poorly because it covers only a few classes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MalCL with FML and L1 distance to per-class mean logits for replay selection instead of None | MalCL · EMBER, 11 tasks | mean accuracy: 27.5 -> 54.5 | Helped |
| Use MalCL with FML and L1 distance to per-class mean logits for replay selection instead of GR | MalCL · EMBER, 11 tasks | mean accuracy: 26.8 -> 54.5 | Helped |
| Use MalCL with FML and L1 distance to per-class mean logits for replay selection instead of BI-R | MalCL · EMBER, 11 tasks | mean accuracy: 27.0 -> 54.5 | Helped |
| Add FML instead of BCE (ablation) | MalCL with L1 to CMean Logits replay selection · EMBER, 11 tasks | mean accuracy: 52.1 -> 54.5 | Helped |
| Add FML instead of BCE (ablation) | MalCL with L1 to CMean Logits replay selection · EMBER, 11 tasks | minimum accuracy: 26.2 -> 21.8 | Hurt |
| Add FML instead of BCE (ablation) | MalCL with L2 to Labels replay selection · EMBER, 11 tasks | mean accuracy: 47.8 -> 50.2 | Helped |
| Add FML instead of BCE (ablation) | MalCL with L2 to Labels replay selection · EMBER, 11 tasks | minimum accuracy: 19.4 -> 17.8 | Hurt |
| Add FML instead of BCE (ablation) | MalCL with L1 to BMean Logits replay selection · EMBER, 11 tasks | mean accuracy: 47.6 -> 50.3 | Helped |
| Add FML instead of BCE (ablation) | MalCL with L1 to BMean Logits replay selection · EMBER, 11 tasks | minimum accuracy: 24.2 -> 32.0 | Helped |
| Use per-class mean logits instead of per-batch mean logits for L1 replay selection with FML (ablation) | MalCL · EMBER, 11 tasks | mean accuracy: 50.3 -> 54.5 | Helped |
| Use per-class mean logits instead of one-hot labels for replay selection with FML (ablation) | MalCL · EMBER, 11 tasks | mean accuracy: 50.2 -> 54.5 | Helped |
| Assign larger classes to the initial task (ablation) | MalCL with L1 to CMean Logits and BCE · EMBER, 11 tasks | mean accuracy: 55% -> 74% | Helped |
Surprising findings
- Image-domain generative replay methods GR and BI-R perform about the same as the None baseline on malware classification.
- L1 to per-batch mean logit selection produces synthetic samples for only a few classes, causing poor performance.
Limitations
- Substantial gap remains to the Joint upper bound (54.5% vs 88.7% on EMBER).
- Class imbalance in EMBER weakens retention and accuracy.
- Global selection scheme may not cover all classes, so mode collapse of the generator is not fully resolved.
Experimental setup
Datasets: EMBER 2018 AZ-Class (AndroZoo)
Benchmarks: EMBER AZ-Class
Metrics: mean accuracy minimum accuracy
Baselines: None Joint GR BI-R
Builds on: Generative Replay Brain-Inspired Replay Improved GAN training (feature matching)
Closest papers by meaning
- Memory Replay GANs: Learning to Generate New Categories without Forgetting (NeurIPS 2018 · similarity 0.85)
- Poisoning Generative Replay in Continual Learning to Promote Forgetting (ICML 2023 · similarity 0.85)
- Feature-Space Bayesian Adversarial Learning Improved Malware Detector Robustness (AAAI 2023 · similarity 0.84)
- DDGR: Continual Learning with Deep Diffusion-based Generative Replay (ICML 2023 · similarity 0.84)
- Attack on Prompt: Backdoor Attack in Prompt-Based Continual Learning (AAAI 2025 · similarity 0.83)
- BrainWash: A Poisoning Attack to Forget in Continual Learning (CVPR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.