Network Intrusion Detection Adaptability under Spurious Labels: A Noise-Aware Few-Shot Class-Incremental Learning Framework
Organizations: College of Computer Science, Beijing University of Technology, China · Department of Electrical Engineering and Information Technologies (DIETI), University of Naples Federico II, Italy · IMT School for Advanced Studies, Lucca, Italy
Abstract
Network Intrusion Detection Systems (NIDSs) must incrementally adapt to emerging attacks, even when only a limited amount of labeled network traffic is available. Consequently, Few-Shot Class-Incremental Learning (FSCIL)-based NIDSs have gained growing attention. However, existing FSCIL-based NIDSs are highly vulnerable to label noise in newly-observed attack classes. To address this limitation, we propose NAFN, a noise-aware FSCIL-based NIDS framework that reduces the impact of mislabeled data during incremental adaptation. NAFN consists of two components: (i) Noise Detection, which identifies and removes mislabeled biflows using noise-detection algorithms, and (ii) Feature Sampling, which estimates class-conditional embeddings from clean biflows, calibrates them with base-class statistics, and resamples synthetic features to recover the original biflow count. The framework follows the Rethink Few-Shot (RFS) teacher-student paradigm with lightweight nearest-neighbor or logistic regression classifiers, and various pretraining & memory selection strategies. Experiments on CSE-CIC-IDS2018, Edge-IIoT, and IoT-NID datasets show that NAFN effectively preserves prior knowledge and maintains strong incremental performance even under 50% label noise, consistently improving new-class F1-scores over non-denoised baselines, up to +52%, while being time-efficient and robust in cross-network and high-noise incremental settings. The implementation is publicly available to foster reproducibility.
Figures & tables
| (a) CSE-CIC-IDS2018 | |||||||||
| Metric | Mem+ | NN Classifier | LR Classifier | ||||||
| W/o Pretrain. | Reconstr. | Contras. | Hybrid | W/o Pretrain. | Reconstr. | Contras. | Hybrid | ||
| F1-New | Random | 93.11 2.14 | 92.81 2.15 | 92.02 2.51 | 93.44 2.03 | 90.02 1.91 | 90.53 1.95 | 84.15 2.90 | 90.17 2.38 |
| Rainbow | 90.74 2.36 | 92.96 2.46 | 93.29 2.26 | 93.35 2.22 | 89.70 2.04 | 90.25 1.81 | 90.24 1.68 | 90.64 1.60 | |
| Herding | 90.51 2.14 | 90.31 2.25 | 91.22 2.43 | 90.31 2.23 | 85.31 3.05 | 90.49 1.13 | 89.01 1.61 | 90.26 1.08 | |
| F1-Old | Random | 98.25 0.53 | 98.26 0.51 | 98.03 0.57 | 98.40 0.46 | 95.38 1.75 | 96.00 0.63 | 90.07 2.89 | 95.95 0.66 |
| CSE-CIC-IDS2018 | Edge-IIoT | IoT-NID | |
| Base Training Time ( ) | |||
| W/o Pretrain | 139.41 0.46 | 740.52 1.97 | 456.39 1.92 |
| Reconstruction | 214.95 0.54 | 832.19 2.41 | 490.81 1.27 |
| Contrastive | 1,352.22 155.49 | 4,669.52 969.00 | 1,066.13 167.92 |
| Hybrid | 1,400.56 113.37 | 4,125.98 752.35 | 1,055.94 152.43 |
| Incremental Training Time ( / increment) | |||
| Base Dataset | Incr. Dataset | Label Noise | |||
| 0% | 88.12 | 97.46 | 60.85 | ||
| 50% | 85.07 | 82.26 | 61.83 | ||
| 50% | 88.12 | 95.68 | 64.10 | ||
| 0% | 90.01 | 58.89 | 50.72 | ||
| 50% | 89.68 | 57.21 | 52.22 | ||
| 50% | 93.45 | 63.45 | 56.49 |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | Classifier | Setting | F1-All | F1-New | F1-Old |
| Denoise-only | 94.96% | 87.04% | 97.59% | ||
| Denoise MixUp | 94.98% | 87.00% | 97.65% | ||
| Denoise SMOTE | 95.00% | 86.91% | 97.65% | ||
| Denoise Ours | 95.01% | 87.02% | 97.66% | ||
| Denoise-only | 78.66% | 32.74% | 93.97% | ||
| Denoise MixUp | 92.81% | 82.94% | 96.18% |
| Dataset | Classifier | Step | F1-All | F1-New | F1-Old |
| #1 | 99.12 0.39 | 96.71 1.50 | 99.38 0.31 | ||
| #2 | 99.06 0.37 | 99.86 0.42 | 98.98 0.40 | ||
| #3 | 97.18 0.65 | 85.18 4.07 | 98.27 0.40 | ||
| #1 | 99.08 0.65 | 96.29 1.57 | 99.40 0.26 | ||
| #2 | 99.17 0.36 | 99.95 0.24 | 99.09 0.39 | ||
| #3 | 94.70 0.70 | 76.20 4.03 | 96.39 0.60 |
Network Intrusion Detection
Network intrusion detection identifies malicious or abnormal network activity to help protect computers, connected devices, and critical infrastructure.
See all 184 papers