Local Layer-wise Differential Privacy in Federated Learning

Yunbo Li, Jiaping Gui, Fanchao Meng, Yue Wu

TL;DR

LaDP-FL introduces a layer-wise adaptive noise injection framework for Local Differential Privacy (LDP) in Federated Learning. By dynamically scaling noise based on layer importance and privacy sensitivity (measured via KL divergence), it slashes average noise injection by 46% while doubling model accuracy compared to standard uniform noise approaches.

What Problem Does This Paper Address?

Federated Learning (FL) allows multiple clients to collaboratively train a machine learning model without directly sharing their private data. However, FL isn't perfectly secure: malicious or "honest-but-curious" participants can still reconstruct private training images or infer dataset membership by analyzing the model updates (gradients or weights) sent to the server.

To thwart these attacks, researchers use Local Differential Privacy (LDP), which injects mathematical noise into local updates before they leave the client's device. The catch? Traditional LDP approaches apply noise uniformly across the entire model. This "one-size-fits-all" strategy introduces excessive noise into layers that are critical for accuracy but contain little private information, severely degrading the final model's performance. LaDP-FL aims to fix this suboptimal privacy-utility tradeoff.

Fig. 1. An FL model with an honest-but-curious client trying to infer private information from the global model. “Gen.” and “Disc.” denote Generator and Discriminator, respectively.

Methods

LaDP-FL replaces monolithic model-level noise with a fine-grained, layer-wise approach consisting of three core modules executed on the client side:

  1. Layer Selection Module: Not all neural network layers contribute equally to performance. Borrowing insights from model pruning, LaDP-FL calculates the $L_2$ norm of each layer's weights. Layers with weights exceeding a threshold $R$ are deemed crucial for prediction and selected for protection. Layers with small weights are skipped, preventing unnecessary noise injection that would harm utility.
  2. Privacy Estimation Module: To determine how much noise a selected layer needs, the framework calculates the Kullback-Leibler (KL) divergence between the local layer's distribution and the global model's layer distribution.
  3. Adaptive Noise Injection Module: Gaussian noise is sampled and injected into each layer. The noise volume is scaled inversely to the KL divergence estimate: layers that closely match the global distribution (low KL divergence) represent a higher privacy risk and receive more noise, while highly distinct layers (high KL divergence) receive less.
Fig. 2. A visual representation of each module of LaDP-FL. The blue and red squares represent different levels of privacy information associated with different layers, resulting in different amounts of noise injection.

Datasets

The authors evaluated LaDP-FL using standard computer vision architectures (CNN and ResNet-18/ResNet-152) across three prominent datasets:

  • CIFAR-10
  • CIFAR-100
  • FEMNIST

To simulate harsh real-world conditions, they tested the framework under pathological non-IID (independent and identically distributed) data setups, including Label Scarcity (where clients only hold a fraction of the total labels) and Quantity Skew (modeled via a Hetero-Dirichlet Distribution).

Key Insights and Ideas

  • Privacy Risk Correlates with Distributional Similarity: A key "aha moment" is using KL divergence as a data-agnostic proxy for privacy risk. If a client's local layer weights look almost identical to the global model's weights, it indicates a high risk of leakage because adversaries rely on stable global properties to train shadow models for reconstruction attacks.
  • Selective Noise Pruning: Skipping noise injection on low-weight layers protects model utility without exposing the client to threats. Because these layers have minimal predictive power, adversaries cannot easily exploit them to build effective shadow models.
Illustration of KL divergence's relationship with privacy risk.

Key Results

  • Massive Accuracy Gains: Across all tested scenarios, LaDP-FL achieved an average accuracy improvement of 102.99% over state-of-the-art (SOTA) DP-FL methods. Under tight privacy budgets ($\epsilon = 0.2$), it outperformed standard Full DP by 33.37% on CIFAR-10.
  • Reduced Noise Footprint: The framework reduced the total volume of injected noise by 46.14% on average compared to SOTA baselines like AdapLDP and Dynamic Privacy Allocation LDP.
  • Robust Attack Defense: When facing generative adversarial network (GAN) reconstruction attacks, LaDP-FL successfully distorted the reconstructed images, increasing the Fréchet Inception Distance (FID) by >12.84% compared to baselines, making the stolen data semantically incoherent to attackers.
  • Resilience to Heterogeneity: In extreme non-IID settings, LaDP-FL maintained a 25.21% accuracy advantage over baseline methods due to its dynamic per-layer noise adaptation.
The influence of the privacy budget on global accuracy and loss when using ResNet-18 on different datasets.

Limitations and Future Work

While LaDP-FL provides superior utility, it does introduce a minor computational bottleneck. Calculating layer-wise weight norms and KL divergences adds an average of 13.56% to 17.05% additional training time per round compared to basic Full DP.

For future work, the authors plan to extend this layer-wise adaptive mechanism to asynchronous Federated Learning environments and explore optimization techniques (such as integrating with Parameter-Efficient Fine-Tuning like LoRA) to minimize the computational overhead on massive, deep-tier models.

Summary updated

Abstract

Federated Learning (FL) enables collaborative model training without direct data sharing, yet it remains vulnerable to privacy attacks such as model inversion and membership inference. Existing differential privacy (DP) solutions for FL often inject noise uniformly across the entire model, degrading utility while providing suboptimal privacy-utility tradeoffs. To address this, we propose LaDP, a novel layer-wise adaptive noise injection mechanism for FL that optimizes privacy protection while preserving model accuracy. LaDP leverages two key insights: (1) neural network layers contribute unevenly to model utility, and (2) layer-wise privacy leakage can be quantified via KL divergence between local and global model distributions. LaDP dynamically injects noise into selected layers based on their privacy sensitivity and importance to model performance. We provide a rigorous theoretical analysis, proving that LaDP satisfies $(ε, δ)$-DP guarantees and converges under bounded noise. Extensive experiments on CIFAR-10/100 datasets demonstrate that LaDP reduces noise injection by 46.14% on average compared to state-of-the-art (SOTA) methods while improving accuracy by 102.99%. Under the same privacy budget, LaDP outperforms SOTA solutions like Dynamic Privacy Allocation LDP and AdapLDP by 25.18% and 6.1% in accuracy, respectively. Additionally, LaDP robustly defends against reconstruction attacks, increasing the FID of the reconstructed private data by $>$12.84% compared to all baselines. Our work advances the practical deployment of privacy-preserving FL with minimal utility loss.
BibTeX
@misc{li2026local,
  author = {Yunbo Li and Jiaping Gui and Fanchao Meng and Yue Wu},
  title = {{Local Layer-wise Differential Privacy in Federated Learning}},
  year = {2026},
  eprint = {2601.01737},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.01737},
}
Figure