TS-GAN: A Lightweight Cross-Domain Adaptation Framework for Network Intrusion Detection

A 92.4% reduction in inference latency was achieved by a student model with only 4.8 million parameters, enabling real-time processing. The framework introduces TS-GAN, a temporally-aware generative adversarial network, to align feature distributions between source and target domains for cross-domain adaptation. In a transfer task from UNSW-NB15 to CIC-IDS2017, the model achieved a 90.13% F1-score and demonstrated resilience under 15% feature perturbation.

Use Cases

• Training lightweight intrusion detection models based on knowledge distillation from a high-performance teacher network.
• Evaluating cross-domain adaptation performance using transfer tasks between datasets like UNSW-NB15 and CIC-IDS2017.
• Conducting robustness stress tests on detection models under controlled feature perturbation.
• Studying the co-evolutionary trend between generation quality (FID) and detection performance (F1-score) in adversarial networks.

Strengths

• Achieved a 92.4% reduction in inference latency, bringing average latency to 1.42 ms per sample.
• Model demonstrated a 90.13% F1-score in a cross-domain transfer task and resilience under 15% feature perturbation.
• Dataset is openly licensed under CC-BY-4.0, facilitating reuse and distribution.

Limitations

• The specific column names and row count for the underlying data are not provided, limiting understanding of its structure.
• The dataset's size is listed as 1012190591 bytes, but the nature of the files (e.g., raw network data, extracted features, model weights) is not explicitly detailed.
• The last update date is listed as 2026-03-20, which is a future date, indicating a potential metadata error.

Provenance

Source

HongGang Miao

Collection Method

Proposed and implemented as part of a research study on lightweight cross-domain adaptation for intrusion detection.

Freshness

2026-03-20