analiza aiqbench
This research introduces DP-DT, a novel framework that decouples representation learning from privacy enforcement to overcome the utility bottlenecks inherent in standard differentially private methods. By confining noise injection solely to weight aggregation and leveraging auxiliary models for feature extraction, the approach significantly improves performance on vision and language benchmarks while maintaining rigorous theoretical guarantees under non-convex objectives. The study demonstrates that this architecture achieves state-of-the-art privacy-utility trade-offs without suffering from iterative noise accumulation or data inefficiency associated with disjoint partitioning schemes.
Differentially Private Neural Network Training Under the Hidden State Assumption
arXiv:2407.08233v3 Announce Type: replace Abstract: Current differentially private learning paradigms face a severe utility bottleneck: DP-SGD degrades performance through noise accumulation over training steps, while aggregation-based approaches such as PATE suffer from data inefficiency due to disjoint data partitioning. We propose \textbf{Differentially Private Decoupled Training (DP-DT)}, a framework that decouples representation learning from privacy enforcement. DP-DT confines noise injection to the weight aggregation stage and employs auxiliary models, continuously synchronized with a global model, to perform noise-free feature extraction on private data shards. We prove that DP-DT converges globally to a limit point under non-convex objectives, via a Lyapunov potential analysis combined with the Kurdyka--\L{}ojasiewicz property. Under the Hidden State Assumption, where adversaries observe only the final published model, we further prove that DP-DT's privacy loss can converge to a constant bound rather than accumulating with iterations. Empirical results across vision and language benchmarks confirm that DP-DT significantly mitigates utility degradation, achieving state-of-the-art privacy-utility trade-offs.
Abstract:Current differentially private learning paradigms face a severe utility bottleneck: DP-SGD degrades performance through noise accumulation over training steps, while aggregation-based approaches such as PATE suffer from data inefficiency due to disjoint data partitioning. We propose \textbf{Differentially Private Decoupled Training (DP-DT)}, a framework that decouples representation learning from privacy enforcement. DP-DT confines noise injection to the weight aggregation stage and employs auxiliary models, continuously synchronized with a global model, to perform noise-free feature extraction on private data shards. We prove that DP-DT converges globally to a limit point under non-convex objectives, via a Lyapunov potential analysis combined with the Kurdyka--Łojasiewicz property. Under the Hidden State Assumption, where adversaries observe only the final published model, we further prove that DP-DT's privacy loss can converge to a constant bound rather than accumulating with iterations. Empirical results across vision and language benchmarks confirm that DP-DT significantly mitigates utility degradation, achieving state-of-the-art privacy-utility trade-offs.
Machine Learning (cs.LG)
arXiv:2407.08233 [cs.LG]
(or arXiv:2407.08233v3 [cs.LG] for this version)
Text integral sub licența deschisă a publicației · 161 cuvinte
Subiecte conexe
Convex Approximation of Two-Layer ReLU Networks for Hidden State Differential Privacy
arXiv · 16 iun. 2026
[AINews] Agents disprove Polysemanticity; greatly improve Interpretability & Alignment
Latent Space · 16 oct. 2026
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · 18 sept. 2026
News
Quantum Computing Report · 25 iul. 2026