Lune

CVPR Findings2026

PLR-Gate: Real-Time Gradient Privacy Assessment and Gated Transmission for Secure Federated Learning

Tao Huang, Jiayang Meng, Hong Chen, Chen Hou, Guolong Zheng, Xu Yang

Abstract

Adversarial Training (AT) has been widely adopted locally in federated learning to enhance robustness against adversarial perturbations. However, adversarial gradients are more vulnerable to gradient inversion attacks and are thus commonly protected by DP-SGD. The dynamic privacy risk of gradients in Differentially Private-AT (DP-AT) is challenging to quantify in real-time using existing static assessments, such as attack-based methods, due to their computational intensity. To address this limitation, we propose PLR-Gate, a novel framework for real-time gradient privacy risk assessment and gated gradient transmission under local DP-AT: (i) Before DP-AT, a Mutual Information Neural Estimator (MINE) is trained on strategically simulated distributions that approximate gradient behavior during AT. This enables real-time estimation of Mutual Information (MI) between gradients and original training images; (ii) Then, based on MI, we define a unified metric-Privacy Leakage Ratio (PLR)-to quantify gradient privacy risk. During DP-AT, PLR-Gate implements a PLR-gated mechanism that selectively transmits gradients according to their estimated PLR values, ensuring adaptive and secure gradient aggregation. Extensive experiments demonstrate that PLR effectively quantifies gradient inversion risk while achieving significantly higher computational efficiency-requiring only about 1% of the runtime of attack-simulation approaches.

Ask Lune

Know this paper in minutes, not an afternoon.

Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.

Questions to start from

Lune can draw on

the full paper · related work

Get the answer

Free to start. No credit card required.

Lineage

  • Bounding training data reconstruction in DP-SGD cites Inverting Gradients - How easy is it to break privacy in federated learning?
  • Bounding training data reconstruction in DP-SGD cites Deep Learning with Differential Privacy
  • Improving Robustness to Model Inversion Attacks via Mutual Information Regularization cites Deep Learning with Differential Privacy
Builds onCited byLeft to right by year; the more cited, the higher.

Builds on9

  1. Deep Learning with Differential PrivacyMartín Abadi, Andy Chu, Ian J. Goodfellow, H. Brendan McMahan et al.CCS 2016 · 7,620 citations
  2. Inverting Gradients - How easy is it to break privacy in federated learning?Jonas Geiping, Hartmut Bauermeister, Hannah Dröge, Michael MoellerNeurIPS 2020 · 1,822 citations
  3. On Adaptive Attacks to Adversarial Example DefensesFlorian Tramèr, Nicholas Carlini, Wieland Brendel, Aleksander MadryNeurIPS 2020 · 1,026 citations
  4. Improving Robustness to Model Inversion Attacks via Mutual Information RegularizationTianhao Wang, Yuheng Zhang, Ruoxi JiaAAAI 2021 · 98 citations
  5. Bounding training data reconstruction in DP-SGDJamie Hayes, Borja Balle, Saeed MahloujifarNeurIPS 2023 · 73 citations

Related papers

  1. Protect Privacy from Gradient Leakage Attack in Federated LearningJunxiao Wang, Song Guo, Xin Xie, Heng QiINFOCOM 2022 · 82 citations
  2. Enhancing Privacy Preservation in Federated Learning via Learning Rate PerturbationGuangnian Wan, Haitao Du, Xuejing Yuan, Jun Yang et al.ICCV 2023 · 2 citations
  3. Surrogate Model Extension (SME): A Fast and Accurate Weight Update Attack on Federated LearningJunyi Zhu, Ruicong Yao, Matthew B. BlaschkoICML 2023 · 17 citations
  4. DictPFL: Efficient and Private Federated Learning on Encrypted GradientsJiaqi Xue, Mayank Kumar, Yuzhang Shang, Shangqian Gao et al.NeurIPS 2025 · 4 citations
  5. Soteria: Provable Defense Against Privacy Leakage in Federated Learning From Representation PerspectiveJingwei Sun, Ang Li, Binghui Wang, Huanrui Yang et al.CVPR 2021
Dusk over the sea between two cliffs drawn in fine vertical lines