Traffic cameras provide useful analysis yet can expose private information. Replacing images with text reduces direct visual exposure but lacks a formal disclosure bound.
The existing reinforcement‑learning plus mixture‑of‑experts (RL+MoE) framework refines scene descriptions using relevance, coverage and brevity rewards, yet it ignores repeated releases under a fixed privacy budget.
This paper extends that framework to budgeted semantic release. Four scene experts first process isolated video chunks and map their outputs to a fixed, bounded semantic schema.
A reinforcement‑learning controller selects which expert summaries to emit and how much privacy budget to spend, based on public task requests and prior private releases.
A separate privacy filter enforces the total budget and adds calibrated noise before converting summaries to text. Under chunk isolation and bounded‑event assumptions, the full release transcript satisfies an event‑level differential privacy guarantee.
The design also permits reuse of earlier summaries without incurring new disclosure.
The evaluation protocol measures traffic utility, privacy attacks, release freshness and computational cost. Illustrative tables show the intended reporting format but do not demonstrate empirical gains.
The main contribution is a task‑aware release interface with auditable privacy accounting for repeated traffic descriptions.
Review