Source-linked AI summary
Few-Shot Learning for Network Intrusion Detection: Methods, Datasets, and Performance
Arne Roszeitis, Victor Jüttner, Erik Buchmann
TL;DR
Few-shot learning addresses the scarcity of labeled attack data that complicates learning-based network intrusion detection. This systematic review analyzes 21 studies from 1,358 records to characterize techniques, datasets, and evaluation settings, finding a field dominated by CNNs and meta-learning but constrained by heterogeneous reporting and limited reproducibility.
Problem
Learning-based NIDS require labeled samples for each attack class, but labels for novel attacks are scarce, motivating few-shot learning from small labeled sets.
Method
The study systematically reviewed 2022–2026 few-shot NIDS research retrieved from ACM Digital Library, IEEE Xplore, and Scopus, retaining 21 studies after screening and quality filtering.
Results
CNNs and meta-learning dominate the reviewed field, appearing in 10 and 8 studies, respectively, while CIC-IDS2017 and CSE-CIC-IDS2018 are the most widely used datasets.
Takeaways & Limitations
Shared evaluation protocols, consistent parameter reporting, open code, and unified benchmarks are identified as priorities for assessing performance and generalization.
Takeaways & Limitations
Missing parameters and unavailable source code prevented verification of reported numbers, so comparisons rely on self-reported results from heterogeneous settings.
Abstract
from arXiv · showhide
Anomaly-based network intrusion detection systems (NIDS) are an important first line of defense. However, training NIDS for new attack types is challenging, because labeled attack data are rarely available. Few-shot learning (FSL) addresses this problem by learning from few samples. However, the approaches and evaluation settings, that have been investigated so far, vary widely. This work systematically reviews FSL approaches for NIDS published from 2022 to 2026. We conduct a systematic literature review with PRISMA 2020-like reporting to search ACM Digital Library, IEEE Xplore, and Scopus. From a set of 1,358 initial records, we retain 21 studies after screening, deduplication, and quality filtering. We classify the applied FSL approaches, datasets, and experimental parameters and compare reported performance. Meta-learning and convolutional neural networks are the most common approaches, with 8 and 10 studies, respectively. Most studies evaluate five or fewer samples per class, although settings vary. CIC-IDS2017 and CSE-CIC-IDS2018 are the most frequently used datasets. Missing parameters and source code limit reproducibility and direct comparison between approaches.
1 Introduction
Few-shot learning offers a way to train NIDS for novel attacks despite scarce labels, but inconsistent techniques and evaluation settings prevent clear comparisons.
- Few-shot learning addresses scarce labeled attack data by learning new intrusion classes from only a small number of labeled samples.
- Existing studies lack consensus on the best FSL techniques and evaluation practices, with “few” ranging from one to 20 samples per class.
- The review asks which learning techniques, datasets, and evaluation settings are used for few-shot NIDS.
- The authors systematically review peer-reviewed studies from 2022 to 2026 across ACM Digital Library, IEEE Xplore, and Scopus, retaining 21 of 1,358 records.
- CNNs and meta-learning are the most common approaches, while CIC-IDS2017 and CSE-CIC-IDS2018 dominate datasets and heterogeneous reporting limits reproducibility.
2 Related Work
Earlier surveys addressed intrusion detection or few-shot learning broadly, but none systematically reviewed post-2021 few-shot learning for network intrusion detection.
- Existing intrusion-detection surveys generally emphasize deep-learning architectures, datasets, or broad NIDS taxonomies rather than few-shot learning.
- General FSL surveys classify techniques and distinguish few-shot learning from related areas such as transfer learning and meta-learning.
- Duan et al. reviewed few-shot intrusion detection through 2021, while Winiecki et al. evaluated selected FSL methods against classical baselines.
- This work addresses the gap by systematically reviewing few-shot NIDS studies published from 2022 to 2026, focusing on techniques, datasets, and evaluation settings.
3 Methodology
The review searches three scholarly databases, applies automated and manual eligibility screening, and extracts techniques, datasets, evaluation settings, and reported performance from included studies.
- The search covered ACM Digital Library, IEEE Xplore, and Scopus using intrusion-detection and few-shot-learning terms, including one-shot learning.
- The screening pipeline reduced 1,358 records to 162 keyword matches, 124 deduplicated studies, 34 quality-filtered studies, and 21 included studies.
- For each included study, the review extracted learning techniques, datasets, class count n, samples per class k, and reported performance where available.
- When multiple settings were reported, the review used k=5 where available or otherwise selected each study’s best-performing setting, with results truncated to three decimals.
- The review groups studies into technique categories derived from the corpus, merging synonymous and subordinate techniques into ten non-exclusive categories.
- Table 2 summarizes publication year, dataset counts, and combined learning techniques for the 21 included studies.
4 Review Results
The review maps FSL techniques, datasets, evaluation settings, and reported performance across 21 NIDS studies, finding recurring method and dataset choices alongside substantial comparison and reproducibility limits.
- 4.1 Overview: CNNs appeared in 10 studies and meta-learning in 8, while 17 of 21 studies combined multiple techniques; their combination occurred in 5 studies.Meta-learning was more common than data-generation approaches, and CNNs appeared throughout the reviewed period.
- 4.2 Datasets: CIC-IDS2017 and CSE-CIC-IDS2018 were used 11 and 6 times, respectively, but many single-use datasets prevented cross-study comparison.Each study evaluated a median of three datasets.
- 4.3 Evaluation and Performance: Among reported evaluation settings, k=5 was most common, 12 of 16 studies specifying k used five or fewer samples per class, and class counts ranged from 2 to 18.Five studies omitted k and five omitted n, limiting reproducibility and comparability.
- 4.3 Evaluation and Performance: Reported best F1-scores ranged from 0.700 to 0.999, with 12 of 15 studies reaching at least 0.900 on one dataset.On CIC-IDS2017, reported F1-scores ranged from 0.673 to 0.995, illustrating variation even on a shared dataset.
- 4.4 Technique-Level Analysis: Technique-level studies included meta-learning, CNN feature extraction, graph neural networks, FSCIL, auto-encoders, synthetic-data generation, metric fusion, and an LLM-based workflow.The technique categories were neither exclusive nor exhaustive, and most studies combined several FSL techniques.
- 4.4 Technique-Level Analysis: Graph neural networks achieved the review’s best reported precision and F1-score, both 0.996 on IoT-23, while CNN results were mixed but often strong on popular datasets.The review reports that CNN approaches were overall better than meta-learning approaches, while noting mixed CNN results.
5 Discussion
The review finds a fragmented FSL-for-NIDS field dominated by CNNs and meta-learning, while reported performance remains difficult to compare because techniques, datasets, and evaluation settings vary. It therefore supports unified protocols, reproducibility practices, and cross-dataset benchmarking.
- 5.1 Interpretation and Implications: Meta-learning and CNNs dominate the reviewed field, but graph neural networks and Auto-Encoder combinations achieve the highest reported scores.The review also observes that widespread technique adoption does not necessarily correspond to the strongest reported results.
- 5.2 Limitations: A venue-based quality filter may have excluded relevant studies from newer or lower-ranked venues because citation-based metrics vary across research communities.The filter required a journal h4-index of at least 53 or a conference ranking of B or better in CORE.
- 5.2 Limitations: Single-researcher screening, eligibility decisions, extraction, and query design leave the review subject to avoidable subjectivity.The authors state that a second reviewer would have reduced this subjectivity.
- 5.2 Limitations: Incomplete parameter reporting, unavailable source code, and heterogeneous settings prevent direct verification and reliable comparison of reported results.These gaps include omitted k or n values and limited access to implementation details.
- 5.3 Future Work: Future work should reimplement representative methods under fixed n and k, common datasets, identical preprocessing, complete metrics, and cross-dataset evaluation.The review notes that none of the included studies trains and evaluates on disjoint datasets, leaving robustness to unseen environments unknown.
6 Conclusion
This review examines recent few-shot learning for NIDS as a response to scarce labeled attack data. Across 21 studies, it finds dominant CNN and meta-learning use but limited comparability and reproducibility, motivating shared, openly reproducible benchmarks.
- 6 Conclusion: The review analyzed 21 peer-reviewed studies selected from 1,358 records published between 2022 and August 2026.It examined learning techniques, datasets, and evaluation settings.
- 6 Conclusion: CNNs and meta-learning dominate the literature, while CIC-IDS2017 and CSE-CIC-IDS2018 serve as de facto benchmarks.The benchmark datasets provide only a partial basis for comparison because evaluation settings remain heterogeneous.
- 6 Conclusion: Heterogeneous settings, incomplete parameter reporting, and largely unavailable source code limit direct comparison and independent verification.These limitations affect interpretation of the reported performance across studies.
- 6 Conclusion: The field needs fixed n-way/k-shot protocols, common datasets, consistent metrics, disclosed code and hyperparameters, and ideally cross-dataset evaluation.The authors identify a unified, openly reproducible benchmark as the most valuable next step for assessing performance and generalization.
Appendix
The appendix contains database-specific search-query information and a table of datasets, parameters, and results. The results table marks averaged, balanced-accuracy, and macro-averaged values with distinct symbols.
- Appendix: Table 4 presents the database-specific search queries used in the review.
- Appendix: Table 5 organizes datasets, experimental parameters, and reported results, with symbols distinguishing averaged and score-type entries.Asterisks mark results averaged over all datasets, double asterisks mark balanced accuracy, and “M” indicates macro-averaged scores.
- Appendix: The appendix table includes fields for year, author, dataset, n, k, accuracy, precision, recall, and F1-score.