PrivEdge-VLM: Risk-Adaptive Privacy-Preserving Edge Vision–Language Analytics for UAV-Assisted Cyber–Physical–Social Systems
Open Access
Abstract
UAV-assisted cyber–physical–social systems increasingly use visual analytics for disaster response, infrastructure monitoring, traffic management, and safety-oriented situational awareness. Raw aerial imagery can expose faces, license plates, private-property context, location traces, and sensitive human activity, while cloud-based vision–language processing can increase latency, bandwidth use, and governance risk. This article presents PrivEdge-VLM, a risk-adaptive privacy-preserving edge vision–language analytics framework for UAV-assisted cyber–physical–social systems. PrivEdge-VLM integrates object detection, privacy-sensitive region detection, mission-aware semantic tokenization, local inference, sanitized split inference, federated LoRA adaptation with differential privacy and simulated masked aggregation, and deterministic operator guardrails. Across public UAV imagery, public privacy benchmarks, and non-human staged object protocol examples, the evaluation compares raw-cloud, downsampling, static-blur, local-VLM, raw-embedding split, federated static-privacy, no-DP, and full configurations. The full configuration preserves mission utility while reducing residual visual PII, plate-OCR leakage, membership-inference risk, embedding inversion, bandwidth, and unsafe identity-oriented responses under the stated attack protocols. The edge hardware deployment separates privacy-front-end throughput from event-level VLM query latency. The results indicate that UAV VLM privacy is a CPS co-design problem involving semantic transformation, attack-based leakage evaluation, resource-aware deployment, and operator oversight.Keywords
This publication has 38 references indexed in Scilit:
- MAVREN: A multilayered adaptive framework for deploying vision-language models on resource-constrained unmanned aerial vehicles for autonomous search and rescueEngineering Applications of Artificial Intelligence, 2025
- Efficient GPT-4V level multimodal large language model for deployment on edge devicesNature Communications, 2025
- Prompt injection attacks on vision language models in oncologyNature Communications, 2025
- VizWiz-Priv: A Dataset for Recognizing the Presence and Purpose of Private Visual Information in Images Taken by Blind People2019
- The Unmanned Aerial Vehicle Benchmark: Object Detection and Tracking2018
- DOTA: A Large-Scale Dataset for Object Detection in Aerial Images2018
- Rényi Differential Privacy2017
- Faceless Person Recognition: Privacy Implications in Social Media2016
- The regulation of civilian drones' impacts on behavioural privacyComputer Law & Security Review, 2014
- Unmanned aircraft systems: Surveillance, ethics and privacy in civil applicationsComputer Law & Security Review, 2012