PrivacyBench: Privacy Isn't Free in Hybrid Privacy-Preserving Vision Systems

A new benchmarking framework called PrivacyBench reveals that combining privacy-preserving techniques in machine-vision systems can cause catastrophic accuracy failures, challenging the assumption that such methods can be layered without consequence [1][2]. The framework, introduced in a paper posted to the arXiv preprint repository, systematically evaluates hybrid configurations of federated learning, differential privacy, and secure multi-party computation across medical imaging datasets [1][2]. The authors found that pairing federated learning with differential privacy caused model accuracy to collapse from 98% to 13% while substantially increasing compute costs and energy consumption [1][2]. In contrast, combining federated learning with secure multi-party computation maintained near-baseline performance with only modest overhead [1][2]. The paper was submitted by researcher Steven Kolawole on February 21, 2026, as a 125 KB manuscript and revised on June 24, 2026, to 130 KB [1]. The experiments were conducted using ResNet18 and Vision Transformer models on medical data [1][2]. The framework automates configuration through YAML files, resource monitoring, and reproducible protocols, providing what the authors describe as the first systematic platform for evaluating privacy-utility-cost trade-offs [1][2]. arXiv, which hosts the paper, is an open-access repository of electronic preprints that has grown to a submission rate of about 24,000 articles per month as of November 2024 [8]. The repository, launched in 1991, surpassed two million articles by the end of 2021 and operates with moderation but without peer review [8]. The PrivacyBench paper appears under the Computer Science category on arXiv, specifically within the Cryptography and Security subfield [1]. The findings indicate that privacy techniques cannot be composed arbitrarily and carry implications for deployment in resource-constrained environments such as autonomous systems and medical imaging [1][2].