Data Privacy in AI/ML: From Secure Systems to Interpretable and Robust Models
Latest 8 papers on data privacy: Oct. 17, 2026
The quest for powerful yet private AI/ML systems has never been more urgent. As AI permeates sensitive domains like healthcare, finance, and autonomous systems, ensuring the confidentiality of data and models, while also guaranteeing robustness and interpretability, presents a formidable challenge. Recent research has been pushing the boundaries, delivering significant breakthroughs in secure multi-party computation (MPC), federated learning, and the fundamental understanding of privacy-robustness tradeoffs. This digest dives into some of the latest advancements, showcasing innovative solutions that promise a more secure and trustworthy AI future.
The Big Idea(s) & Core Innovations
One of the most profound overarching themes is the drive towards scalable, privacy-preserving computation. Traditional centralized AI models often require pooling vast amounts of sensitive data, creating privacy risks. Several papers tackle this by leveraging techniques that allow models to learn from decentralized data without direct exposure. For instance, NOMAI: Scalable Federated Learning with Private Weights by Yiyi Sun, Mushtari Sadia, and Amrita Roy Chowdhury from the University of Michigan, Ann Arbor, introduces a groundbreaking MPC-based federated learning (FL) system. NOMAI is the first to simultaneously offer data privacy, weight privacy, and integrity guarantees for millions of users, achieving a three orders-of-magnitude improvement over prior work. This is made possible by a novel multi-committee MPC architecture combined with packed Shamir secret sharing and ML-aware optimizations like Winograd convolution. Their core insight is that packed Shamir secret sharing enables SIMD-style secure computation that scales dramatically beyond monolithic n-party protocols.
Building on the strength of federated learning, DeFiRank: Federated Pairwise Ranking for Governance in Decentralized Financial Networks by Zahra Batool et al. from École de technologie supérieure, Montréal, Canada, demonstrates how FL can enable multiple institutions to collaboratively train a ranking model for DeFi governance without sharing private transaction data. Their LSTM-based TemporalRankNet achieves ranking stability comparable to centralized approaches while preserving privacy, showcasing FL’s potential beyond traditional model training. Similarly, for critical infrastructure, SecFD-FL: Frequency-Decoupled Federated Learning with Secure Aggregation for Charging Load Forecasting in Internet of Electric Vehicles by Mamunur Rashid et al. from Tennessee Technological University and University of Ottawa, introduces a privacy-preserving FL framework for EV charging load forecasting. This system employs CEEMDAN-driven frequency decomposition and a lightweight zero-sum additive masking protocol, achieving high accuracy and robust privacy against gradient inversion attacks—a significant leap in securing volatile time-series data.
Beyond privacy in computation, there’s a strong emphasis on formal verification and interpretability. The paper SMT-Boosted Security Types and Correctness for Low-Level MPC by Christian Skalka et al. from the University of Vermont, USA, develops a new Hoare logic and type theory to automatically enforce correctness, confidentiality, and integrity properties of low-level MPC protocols. Their approach, agnostic to specific secret sharing schemes, uses SMT verification to enforce “noninterference modulo output,
Share this content:
Discover more from SciPapermill
Subscribe to get the latest posts sent to your email.
Post Comment