Poisson subsampling is the default sampling scheme in differentially private machine learning, largely because its unstructured randomness yields tractable privacy amplification analyses. Yet this same randomness introduces substantial participation variance: each sample appears in very different numbers of training iterations. In this work, we show that this variance is not merely a practical artifact to be tolerated, but a fundamental source of suboptimal privacy amplification. We prove that Balanced Iteration Subsampling (BIS), a structured scheme in which each sample participates in exactly a fixed number of iterations, achieves stronger privacy amplification than Poisson subsampling and is optimal at both extremes of the noise spectrum ($σ\to 0$ and $σ\to \infty$). Our analysis reveals that the privacy-noise tradeoff is governed not by maximizing randomness, but by eliminating participation variance while preserving uniform marginal participation across iterations. To translate this asymptotic theory into finite-noise guarantees, we introduce a practical near-exact Monte Carlo accountant for BIS, which removes the analytical slack of existing RDP and composition-based PLD analyses. Evaluations across more than 60 practical DP-SGD configurations show that BIS consistently outperforms Poisson subsampling in the low-noise regimes most relevant for high-utility private training, reducing the required noise multiplier by up to $9.6\%$. These results overturn the common intuition that more sampling randomness necessarily yields stronger privacy amplification: in DP-SGD, structured participation can be both more practical and more private. Our implementation is available at https://github.com/dong-xin-ao-andy/bis-mc-accountant.

March 17, 2022--Moshe Shenfeld, a computer science Ph.D. candidate at Hebrew University of Jerusalem (HU)'s Rachel and Selim Benin School of Engineering and Computer Science, was selected as an Apple Scholar in AI/Machine Learning for 2022. Shenfeld is one of only 15 awardees worldwide, the other Israeli recipient is from Tel Aviv University. The Ph.D. fellowship in Machine Learning and AI was created by Apple "to celebrate the contributions of students pursuing cutting-edge fundamental and applied machine learning research worldwide." Currently, Shenfeld is researching privacy-preserving machine learning under the supervision of HU Professor Katrina Ligett. His Ph.D. focuses on differential privacy and its relation to adaptive data analysis and machine learning.

Brakes squeal, hard drives crunch, air conditioners rattle, and their owners know it's time for a service call. But some of the most valuable machinery in the world often operates with nobody around to hear the mechanical breakdowns, from the chillers and pumps that drive big-building climate control systems to the massive turbines at hydroelectric power plants. That's why a number of startups are working to train computers to pick up on changes in the sounds, vibrations, heat emissions, and other signals that machines give off as they're working or failing. The hope is that the computers can catch mechanical failures before they happen, saving on repair costs and reducing downtime. "We're developing an expert mechanic's brain that identifies exactly what is happening to a machine by the way that it sounds," says Amnon Shenfeld, founder and CEO of 3DSignals, a startup based in Kfar Saba, Israel, that is using machine learning to train computers to listen to machinery and diagnose problems at facilities like hydroelectric plants and steel mills.

An innovative AI startup from Israel is using deep-learning AI technology to listen to machinery and predict whether it's about to go wrong. Wondering if your car engine has a problem just got a lot easier. A lot of the most high-profile applications of deep learning technology involve aspects of computer vision, such as cutting-edge facial-recognition technology. However, an innovative artificial intelligence startup from Israel is looking to apply those same neural networks and smart algorithms to another area -- acoustics. Better yet, they are doing so to help users spot early warning signs that machines, such as cars, may be about to fail. "I was on a train about three years ago, going back to my hotel after a business meeting," 3DSignals CEO Amnon Shenfeld told Digital Trends.

A lot of the most high-profile applications of deep learning technology involve aspects of computer vision, such as cutting-edge facial-recognition technology. However, an innovative artificial intelligence startup from Israel is looking to apply those same neural networks and smart algorithms to another area -- acoustics. Better yet, they are doing so to help users spot early warning signs that machines, such as cars, may be about to fail. "I was on a train about three years ago, going back to my hotel after a business meeting," 3DSignals CEO Amnon Shenfeld told Digital Trends. "The train suddenly started making strange noises. These weren't the usual sounds that trains make, but something out of the ordinary. It was the same week that a train had overturned in Spain, injuring a lot of people. Everyone stopped talking and was getting very worried. I had the thought that maybe if there was a train mechanic or engineer sitting on the train, they could tell us if the noise was normal and, if not, where it was coming from and whether it could be safely ignored."