Work

FBBDL: flow-aware benign-behavior deviation learning for leakage-controlled unseen-family DoS detection

Year: 2026
Type: article
Abstract: Detecting previously unseen denial-of-service (DoS) attack families remains challenging because conventional random or stratified splits may allow family-specific information to influence model development, while imbalanced attack distributions can bias detection toward dominant families. This study proposes Flow-Aware Benign-Behavior Deviation Learning (FBBDL), a leakage-controlled framework for unseen-family DoS detection using aggregated network-flow statistics. FBBDL excludes every bin containing a target-family flow from all development stages and integrates duplicate-group-disjoint cross-fitted benign-deviation estimation, tempered family-frequency weighting, nested pseudo-unseen-family validation, and development-only false-positive-rate (FPR) control. Experiments on 692,692 eligible flows from the Wednesday CIC-IDS2017 capture evaluated four independently withheld DoS families over 20 matched seeds. FBBDL achieved a macro recall of 95.64%, a macro geometric mean (G-mean) of 0.975, and a worst-family G-mean of 0.958, with a mean benign FPR of 0.62%. Compared with class-weighted XGBoost (CW-XGB), the strongest evaluated baseline, FBBDL improved macro and worst-family G-mean by 0.016 and 0.033, respectively; the paired differences remained significant after Holm correction (( p < 0.001)). Ablation experiments identified multi-flow temporal aggregation as the dominant contributor, with additional gains from benign-reference deviation, family-frequency weighting, and nested family withholding. Sensitivity analyses further indicated stable performance across the evaluated aggregation and balancing settings. The frozen detector required 0.81 ms per aggregated bin, corresponding to approximately 1,235 bins per second. An independent Wednesday-to-Friday transfer evaluation revealed important limitations: although DDoS score ranking remained strong (normalized pAUROC@1% = 0.997), the frozen threshold produced a benign FPR of 30.23%, exceeding the 1% operating target, while PortScan recall was only 0.14%. These findings support the effectiveness of FBBDL under the evaluated within-capture family-withheld setting, while demonstrating that reliable cross-capture operation cannot be inferred from score-ranking performance alone.
Language: English
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7234781816
FWCI:
Cites: 44
Cited by:
Related to: 10