Cross-Dataset Evaluation Results for Network Intrusion Detection

Cross-dataset evaluation results (S9–S16) from the TAN-IDS framework, a method for assessing NetFlow-based intrusion detection models. The dataset, created by Dung Ha Thanh and shared on figshare in April 2026, contains performance metrics from experiments testing model robustness across different network environments. It is a small dataset at 5.5 KB, stored in an XLS file.

Use Cases

• Analyze cross-dataset performance degradation for models using the 8-dimensional NetFlow feature interface.
• Compare attack-class recall and F1-macro scores across evaluation scenarios like direct transfer or fine-tuning.
• Benchmark lightweight Transformer-based control models against other architectures on domain shift metrics.
• Evaluate the effectiveness of mixed-domain training strategies using reported generalization metrics.
• Study the impact of using 5% labeled target-domain data on model recovery in fine-tuning scenarios.

Strengths

• Formalized evaluation across 4 explicit deployment scenarios (in-dataset, cross-dataset, mixed-domain, fine-tuning).
• Unified data representation using a constrained 8-dimensional NetFlow feature interface.
• Results demonstrate fine-tuning effectiveness, with performance recovery exceeding 95% F1-macro in several cases.

Limitations

• Very small dataset scale (5.5 KB), indicating limited raw experimental data.
• Unknown row count and specific column structure limits detailed reproducibility.
• Results are specific to NetFlow-based IDS and the TAN-IDS framework, limiting generalizability.

Provenance

Source

figshare, authored by Dung Ha Thanh.

Collection Method

Results from experiments within the TAN-IDS evaluation framework using representative ML models.

Time Range

null

Freshness

Last updated April 2026.

Geography

null