跳转至

CutClean：面向隐私保护推理的神经网络剪枝方法

文章背景与核心概要

随着神经网络在各类高风险场景中的广泛部署，模型推理过程中的隐私泄露问题日益受到关注。传统观点往往将隐私泄露归咎于导致传统数据集偏差的表征不平衡，然而本文研究表明，即便在没有此类表征不平衡的情况下，隐私泄露依然可能发生，这为处理敏感属性的模型部署带来了显著的安全隐患。

为了应对这一挑战，本文提出了 CutClean 这一全新的隐私感知神经网络剪枝方法。该方法能够在减少网络中隐私信息流动的同时提高模型的稀疏性。其核心技术在于：在网络的每个区块（block）处引入辅助线性隐私头来量化信息泄露，并通过施加渐进式稀疏性来消除私密属性的泄露（以连接在最后一个区块上的隐私头的准确率为衡量标准）。在合成数据集和真实世界数据集上的实验表明，该方法能够有效最小化私密信息的流动，同时保持高稀疏率并维持目标分类的准确性。

摘要 (Abstract)

Neural networks are increasingly deployed in high-stakes applications with growing privacy leakage concerns. We show that this privacy leakage can occur even in the absence of representation imbalances that lead to traditional dataset biases. This poses significant privacy risks when deploying models that process sensitive attributes.

Neural networks are increasingly deployed in high-stakes applications with growing privacy leakage concerns. We show that this privacy leakage can occur even in the absence of representation imbalances that lead to traditional dataset biases. This poses significant privacy risks when deploying models that process sensitive attributes.

In this context, we propose CutClean, a privacy-aware pruning method that allows to reduce privacy information flow through the network, while increasing its sparsity. Our approach employs auxiliary linear privacy heads placed at each network's block to quantify information leakage, and further applies increasing levels of sparsity to remove the private attribute leakage, measured in terms of the accuracy of the privacy head attached to the last block. Experiments on synthetic and real-world datasets demonstrate that our approach effectively minimizes private information flow while achieving high sparsity rates and preserving classification target accuracy.

In this context, we propose CutClean, a privacy-aware pruning method that allows to reduce privacy information flow through the network, while increasing its sparsity. Our approach employs auxiliary linear privacy heads placed at each network's block to quantify information leakage, and further applies increasing levels of sparsity to remove the private attribute leakage, measured in terms of the accuracy of the privacy head attached to the last block. Experiments on synthetic and real-world datasets demonstrate that our approach effectively minimizes private information flow while achieving high sparsity rates and preserving classification target accuracy.

论文元数据 (Paper Metadata)

  • arXiv ID: arXiv:2608.13773 [cs.LG]
  • 学科分类 (Subjects): 机器学习 (cs.LG); 人工智能 (cs.AI)
  • 作者 (Authors):
  • Leonardo Magliolo
  • Vito Paolo Pastore
  • Giuseppe Valenzise
  • Enzo Tartaglione
  • 提交时间 (Submitted On): 2026年8月13日
  • DOI: 10.48550/arXiv.2608.13773
  • 期刊参考 (Journal Reference): In: De Marsico, M., et al. Pattern Recognition. ICPR 2026. Lecture Notes in Computer Science, vol 16822. Springer, Cham.
  • arXiv ID: arXiv:2608.13773 [cs.LG]
  • Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
  • Authors:
  • Leonardo Magliolo
  • Vito Paolo Pastore
  • Giuseppe Valenzise
  • Enzo Tartaglione
  • Submitted On: 13 August 2026
  • DOI: 10.48550/arXiv.2608.13773
  • Journal Reference: In: De Marsico, M., et al. Pattern Recognition. ICPR 2026. Lecture Notes in Computer Science, vol 16822. Springer, Cham.

访问与资源 (Access & Resources)