Sensor‑intensive environments fuse heterogeneous streams to enable intelligent services, yet users often wish to keep certain activities private. Existing solutions such as differential privacy or rule‑based filters protect individual streams but cannot mitigate cross‑sensor inference risks. PrivateHub addresses this gap by embedding contrastive learning into a diffusion model to synthesize multi‑sensor streams that preserve detectability of non‑private applications while obscuring private ones. The framework consists of two stages:
- App‑Conditioned Pre‑training (ACP), where the model is conditioned on multi‑sensor data together with application embeddings;
- App‑Aware Fine‑tuning (AAF), which employs a contrastive loss to separate private from non‑private samples. We also define a threat model for the multi‑sensor sharing scenario, assuming an adversary can obtain synthetic data and attempt to retrain a classifier. Experiments on three real‑world multi‑sensor datasets show that PrivateHub reduces private‑application accuracy by 40%‑50% without degrading non‑private performance, and remains robust when the attacker retrains on the synthetic data.
Review