Nesa's Academic Papers

Nesa is committed to advancing cutting-edge research at the intersection of privacy, scalability, and trust in AI systems. Our work addresses fundamental challenges in decentralized AI, focusing on secure inference, model partitioning, encrypted computation, and efficient model deployment across heterogeneous environments. Through active publication in peer-reviewed venues, we aim to share innovations that support robust, transparent, and scalable AI infrastructure.

This page presents selected academic papers authored by the Nesa team and collaborators, including accepted conference papers and ongoing preprints. Together, they demonstrate our technical depth across areas such as encrypted model inference, decentralized model sharding, privacy-preserving systems, and meta-learning for deployment optimization.

[1] Model Agnostic Hybrid Sharding for Heterogeneous Distributed Inference

Conference: KDD 2024 — The Fourth International Workshop on Smart Data for Blockchain and Distributed Ledger (SDBD'24) arXiv: arXiv:2407.19775

A hybrid framework addressing privacy, scalability, and heterogeneity in deploying large AI models in decentralized systems.

[2] Towards Secure and Private AI: A Framework for Decentralized Inference

Conference: KDD 2024 — The Fourth International Workshop on Smart Data for Blockchain and Distributed Ledger (SDBD'24) arXiv: arXiv:2407.19401

Presents a secure and scalable framework for decentralized LLM inference, focused on model integrity and privacy.

[3] Meta-Learning for Speeding Up Large Model Inference in Decentralized Environments

Conference: Conference on Language Modeling (COLM) 2025 (Main Conference) arXiv: arXiv:2410.21340

Explores meta-learning techniques to reduce latency and cost for decentralized deployment of large models.

[4] Encrypted Large Model Inference: The Equivariant Encryption Paradigm

Status: Preprint on arXiv arXiv: arXiv:2502.01013

This work introduces a novel encryption paradigm tailored for large-scale models deployed in distributed and decentralized settings. It addresses key privacy vulnerabilities arising from inference-time data exposure and proposes an equivariant encryption approach for enhanced security and performance.