Differentially Private Neural Network Training Under the Hidden State Assumption

We introduce DP-DT, a framework that separates representation learning from privacy protection mechanisms, restricting noise injection to weight aggregation while using auxiliary models for feature extraction on private data. We establish convergence guarantees for non-convex objectives and show that privacy loss can converge to a constant bound rather than accumulating with iterations under the Hidden State Assumption. Empirical validation across vision and language tasks shows improved privacy-utility trade-offs compared to existing methods.