Tag
The paper introduces target-speaker unlearning ASR (TSU-ASR) and proposes a novel Enrollment-Conditioned Gating module for dynamic opt-out of speakers during inference in LLM-based ASR, enhancing privacy in online conferencing.
The paper introduces a federated learning framework for Averaged n-Dependence Estimators (AnDE) classifiers, enhancing privacy by avoiding transmission of semantically meaningful parameters and demonstrating effectiveness on discrete datasets.
Introduces DP-SPIN, a differentially private framework for aggregate measurement and summarization over semantic concepts in text data, with evaluations on CFPB, Amazon, and Yelp datasets.
The paper proves almost sure convergence of stochastic gradient descent with clipping and additive noise, including momentum variants, under smoothness and bounded gradient noise assumptions, providing theoretical foundations for stable training in convex and nonconvex settings.
RAPTOR is a role-aware framework for differentially private fine-tuning of Mixture-of-Experts models, addressing failure modes like clipping interference and noise dilution, with experimental improvements over standard DP baselines.
This paper proposes SAPE-FL, a similarity-aware personalized federated learning framework that adapts to heterogeneous environments by anchoring models to global and peer-averaged models, improving robustness and performance.
Perplexity introduces hybrid compute for its Mac app, enabling local model inference for sensitive data while offloading to the cloud, marking a trend towards transparent local AI usage.
IsleNet introduces a spatial-entropy-based partitioning method for spatiotemporal graph unlearning, enabling exact data removal with low computational cost while maintaining high accuracy for privacy regulations.
The paper investigates the geometric challenges in merging differentially private task models and introduces DP-Merging, a framework to enhance mergeability while maintaining privacy guarantees.
The study reveals substantial gaps in machine unlearning for LLMs, showing that adversarial evaluation uncovers recoverability of forgotten information despite strong standard metrics, highlighting the need for adversarial stress-testing.
This paper studies federated active learning in low-budget regimes, revealing that homogeneous data requires stronger coordination due to heterogeneity reversal. It proposes a framework using federated representation learning to enable globally coordinated active selection, outperforming existing methods.
FedImp is a novel federated learning algorithm that uses impurity-based weighting to enhance convergence speed in non-IID data settings, showing significant reductions in communication rounds compared to baselines.
This paper presents a comprehensive survey of federated prompt learning (FPL), reviewing advances in integrating federated learning with large language models, discussing motivations, trade-offs, and future research directions.
The paper proposes AdaPop, an adaptive popularity-based method for LLM unlearning that adjusts gradient pressure based on fact frequency to improve forgetting effectiveness and reduce leakage under queries.
This paper proposes Hierarchical Federated Transfer Learning (HFTL) for Digital Twin-based Vehicular Ad hoc Networks, addressing data heterogeneity and sparsity via vehicle clustering and a data quality score mechanism to defend against malicious vehicles.
This paper presents the first systematic study of federated training for SpeechLLM-based end-to-end ASR systems, evaluating on English and Italian tasks and achieving competitive word error rates with reduced communication costs.
DP-FedSOFIM moves curvature estimation to the server in differentially private federated learning, achieving the same privacy guarantee as DP-FedGD with O(d) client memory and significant early-round accuracy gains.
The paper proposes OrchNAS, an energy-aware personalized federated edge intelligence framework that uses a Neural Architecture Search service to automatically design service-adaptive models for heterogeneous edge environments, addressing energy constraints and statistical heterogeneity.
OpenWorker is an open-source AI agent that performs tasks like drafting documents, sending messages, and managing calendars, using local data and multi-model support.
Introduces RELIC, a framework for learning interpretable and composable skills in multi-agent planning via revealed principles, enabling privacy-preserving coordination and cross-agent skill transfer without sharing code.