DP SGD
PulseAugur coverage of DP SGD — every cluster mentioning DP SGD across labs, papers, and developer communities, ranked by signal.
8 day(s) with sentiment data
-
New SAVANT model balances privacy and utility in tabular data
Researchers have developed SAVANT, a novel self-attentive variational autoencoder designed for tabular data that integrates differential privacy (DP-SGD). This model is the first to combine attention mechanisms, a VAE, …
-
Apple researchers fix privacy leak in AI training
Apple researchers have identified a privacy vulnerability in the training of private AI models. They demonstrated that observable batch sizes can compromise the privacy guarantees of DP-SGD (Differentially Private Stoch…
-
New DP-SGD method enhances privacy by obscuring batch sizes
Researchers have developed a new mechanism to enhance privacy in Differentially Private Stochastic Gradient Descent (DP-SGD) by privatizing batch sizes. This method uses a random number of dummy points to obscure the ac…
-
New algorithm speeds up privacy loss accounting for random allocation
Researchers have developed a new algorithm called multiscale FFT that significantly speeds up privacy loss accounting for random allocation sampling schemes. This method improves upon existing techniques by reducing the…
-
New Gaussian Approximation Enhances Differentially Private Algorithm Analysis
Researchers have developed a new method for analyzing the privacy guarantees of differentially private algorithms, specifically focusing on differentially private stochastic gradient descent (DP-SGD). The new approach u…
-
Guided Diffusion offers private synthetic text generation
Researchers have introduced Guided Diffusion, a novel framework for generating synthetic text that prioritizes privacy. This method leverages Diffusion Language Models to steer the generation process towards fluent and …
-
New DP-μP framework enables hyperparameter transfer for private model training
Researchers have developed DP-μP, a new framework that extends Maximal Update Parametrization (μP) to differentially private optimization. Standard μP, which allows for zero-shot hyperparameter transfer in normal traini…
-
Research paper examines privacy risks in synthetic data for LLM training
A new research paper explores the trade-offs between model capability and privacy when training on sensitive data, particularly focusing on synthetic medical records. The study found that direct fine-tuning yields the b…
-
Privacy-aware rewriting of synthetic data balances model capability and privacy
A new paper explores the privacy implications of using synthetic data for training language models, particularly with sensitive medical records. The research indicates that privacy-aware rewriting techniques offer a bet…
-
Rényi DP accountants show improved privacy guarantees for DP-SGD
A new research paper explores the effectiveness of Rényi DP accountants for differentially private stochastic gradient descent (DP-SGD). The study observes that Rényi DP can utilize a distinct pair of worst-case distrib…
-
New framework sets privacy-sensitive reconstruction bounds for DP-SGD models
Researchers have developed a new information-theoretic framework to establish reconstruction bounds for models trained with Differential Privacy (DP-SGD). This framework provides a distortion bound that can be calculate…
-
New formula calibrates DP-SGD noise against membership inference attacks
Researchers have developed a new closed-form formula to calibrate noise levels for DP-SGD, a method used to protect training data by adding noise to gradients. This formula specifically applies to DP-SGD with random all…
-
New DP-SGD noise calibration offers tighter bounds against membership inference attacks
Researchers have developed a new, efficient method for calibrating noise in DP-SGD (Differentially Private Stochastic Gradient Descent) training. This approach provides a one-line formula to bound the accuracy of member…
-
Untied embeddings boost private LLM training accuracy and efficiency
A new study published on arXiv explores the impact of weight tying in decoder-only Large Language Models (LLMs) when fine-tuned using Differentially Private Stochastic Gradient Descent (DP-SGD). The research found that …
-
New diffusion models enhance time-series imputation with privacy and global context
Researchers have developed new diffusion model techniques for time-series imputation, focusing on improving accuracy and privacy. One approach, detailed in arXiv cs.LG, uses differentially private diffusion models with …
-
Differential privacy fails to protect vulnerable subgroups in synthetic text releases
A new audit of differentially private synthetic text releases reveals that while differential privacy (DP) is effective at reducing average membership inference leakage, it disproportionately protects some records over …
-
Federated Intrusion Detection Faces Privacy, Robustness, and Fairness Trade-offs
A new research paper explores the trade-offs between privacy, robustness, and fairness in federated learning for network intrusion detection systems (NIDS). The study highlights that while federated learning allows for …
-
New framework enables differentially private synthesis of paired table-image data
Researchers have developed DP-TabImage, a new framework for synthesizing paired tabular and image data while maintaining differential privacy. This approach addresses the challenge of preserving the dependence between m…
-
New framework boosts empirical privacy for DP-SGD in machine learning
Researchers have introduced a new framework to enhance the empirical privacy of machine learning algorithms, specifically targeting DP-SGD. This framework aims to optimize for empirical privacy lower bounds, complementi…
-
New framework enhances privacy in federated driver monitoring
Researchers have developed SecureDrive-FL, a new framework for federated driver monitoring that enhances privacy and security. This system combines Differential Privacy Stochastic Gradient Descent (DP-SGD) with a novel …