PrivLightScreen: a privacy-preserving lightweight multitask AI framework for operational millimeter-wave airport security screening
Year: 2026
Type: article
Abstract: Airport passenger security screening is both a public-safety control and a performance-critical node in the airport supply chain, yet most artificial-intelligence studies on millimeter-wave (MMW) screening optimize detection accuracy without explicitly modeling false alarms, queue delay, privacy risk, resilience, or deployment cost. This paper presents PrivLightScreen, a unified multitask framework for privacy-preserving concealed-object detection in MMW screening, together with an Operational Performance Evaluation Framework (OPEF) that translates inference-level measurements into checkpoint-level performance indicators. PrivLightScreen combines physics-informed adaptive preprocessing, zone-guided self-supervised contrastive pretraining, a lightweight multitask module with an inter-zone Zone Transformer detection head, a representation-level anatomical abstraction layer, and a four-tier knowledge-distillation hierarchy spanning server graphics processing units (GPUs) to embedded processors. On the public Transportation Security Administration (TSA) passenger screening dataset, evaluated with a subject-disjoint protocol, the full model achieves a zone-level F1-score of 0.978 (95% confidence interval (CI) 0.974–0.982), a log-loss of 0.0498 (95% CI 0.0476–0.0521), and an area under the receiver operating characteristic curve (AUC) of 0.983 (95% CI 0.979–0.987), giving the best result among the compared baselines under this protocol. Auxiliary keypoint and segmentation performance is evaluated against weak labels derived from the TSA body-zone metadata and a quality-audited pseudo-mask pipeline, reaching a percentage of correct keypoints normalized by head size ([email protected]) of 99.81 percent and a weak-mask mean intersection over union (mIoU) of 85.7 percent. The smallest distilled student retains 98.3 percent of teacher F1 with a 3.1 MB footprint and an $$8.0\pm 0.6$$ ms batch-size-one forward-pass latency on commodity embedded hardware. Under OPEF, this edge model is the only tested configuration to satisfy all deployment-readiness thresholds, reaching a modeled 353 passengers per hour per lane (95% interval 345–361), a mean resilience index of $$0.82\pm 0.03$$ across six stress conditions, and a design-time Privacy Compliance Score of 0.95. The results show that detection accuracy, operational throughput, resilience, privacy, and hardware efficiency can be evaluated jointly in a single deployable MMW-screening framework, while regulatory deployment still requires site-specific validation and formal certification.
Source: Scientific Reports
Institutions: Air University, George Mason University, University of Dubai
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7225382274