Overview of the proposed semantic smoothing defense. Given a potentially adversarial input, a geometry-conditioned novel view synthesis model generates a set of randomized, semantics-preserving views. Each generated view is evaluated by the classifier, and the final prediction is obtained by aggregating the individual predictions through majority voting。

Overview of the STAMP framework: Tokens are categorized according to their task and privacy relevance, then perturbed using assigned privacy budgets via the polar mechanism. 

• Figure (a) (left) illustrates the overall token perturbation pipeline with group-wise privacy budget allocation; 

• Figure (b) (right) details the grouping and budget assignment process based on task and privacy importance.

We first derive an information-theoretic bound that decomposes the token rejection rate into contributions from SLM-LLM distribution mismatch and from quantization distortion. Guided by this analysis, we propose the Sparse Quantize-and-Sample SD (SQS-SD) framework, which exploits distributional sparsity through structured sparsification and lattice-based quantization.

We develop a framework that combines deep learning and epidemic models to jointly forecast and learn epidemic dynamics using multiple datasets and show that even DP-protected financial data improves forecasting.