LDP-Slicing: Local Differential Privacy for Images via Randomized Bit-Plane Slicing

Yuanming Cao, Chengqi Li, Wenbo He

TL;DR

Local Differential Privacy (LDP) has long been deemed impractical for raw images due to the massive noise injected into high-dimensional pixel spaces. This paper introduces LDP-Slicing, a lightweight, training-free framework that solves this domain mismatch by decomposing pixels into binary bit-planes and optimizing the privacy budget across them. The result is a system that provides rigorous, pixel-level privacy guarantees while maintaining high accuracy for downstream computer vision tasks.

What Problem Does This Paper Address?

Centralized Differential Privacy (DP) protects data by adding noise at the server level, but it requires users to trust a central curator. Local Differential Privacy (LDP) shifts this protection to the data source (the user's device), eliminating the need for a trusted third party.

However, applying standard LDP to images causes a catastrophic drop in utility. Because a standard 8-bit pixel has 256 possible states, traditional LDP mechanisms inject overwhelming noise to make these states statistically indistinguishable, leading to the "curse of dimensionality." Previous attempts bypassed this by applying LDP to low-dimensional feature vectors or embeddings, but this destroys the standard image format, making them incompatible with standard, plug-and-play vision pipelines.

Methods

Figure 1. The LDP-Slicing framework. Our method consists of two primary stages: (1) Perceptual obfuscation: the input image is transformed into the frequency domain via DWT, where the low-frequency (LL) band is pruned to remove human-perceptible information. (2) Bit-plane randomization: The obfuscated image is decomposed into binary bit-planes. A utility-aware randomized response mechanism is applied to each bit and enforces a strict ε-Local Differential Privacy guarantee before the final image is reconstructed.

LDP-Slicing bridges this gap using a three-stage pipeline that operates directly on the image representation without requiring any model training:

  1. Perceptual Obfuscation via Wavelet Pruning: The image is converted to the frequency domain using a 1-level Haar Discrete Wavelet Transform (DWT). The low-frequency approximation sub-band (LL) is completely pruned (set to zero). This removes coarse shapes that humans use to recognize faces, while preserving the high-frequency details that convolutional neural networks (CNNs) leverage.
  2. Bit-Plane Slicing (BPS): Instead of treating a pixel as a single categorical value from 0 to 255, the obfuscated image channels (Y, Cb, Cr) are decomposed into a sequence of binary bits, yielding 24 bit-planes total.
  3. Utility-Aware Privacy Budget Optimization: Rather than distributing the privacy budget ($\varepsilon$) uniformly, the authors use Lagrange multipliers to allocate more budget (less noise) to the Most Significant Bits (MSBs) and the Luma (Y) channel, which carry the bulk of the structural information. Standard binary randomized response is then applied at the bit level before reconstructing the final privatized image.

Datasets

The authors evaluated their framework across multiple domains:

  • Face Recognition: Trained on MS1M-ArcFace (MS1MV2) and evaluated on four standard benchmarks: AgeDB-30, LFW, CPLFW, and CALFW. Cross-dataset zero-shot robustness was tested on VGGFace2 and CelebA.
  • Image Classification: Evaluated using CIFAR-10 and CIFAR-100.
  • Medical Imaging: Tested on the Chest X-Ray dataset to demonstrate domain generalizability.
  • Reconstruction Attacks: Inversion models were trained on the public BUPT dataset.

Key Insights and Ideas

  • Domain Mismatch, Not LDP Limitation: The severe utility loss in image LDP isn't an inherent flaw of the privacy model itself, but rather a consequence of applying it to an improper data representation (categorical pixel values vs. binary bit-planes).
  • Asymmetric Perception: Humans and CNNs "see" differently. Pruning low frequencies completely blinds human observers while leaving enough high-frequency texture for deep learning models to achieve near-baseline performance.
Figure 2. Bit-plane slicing reveals the non-uniform distribu- tion of structural information. An 8-bit image (left) is decom- posed into its planes (right), from LSB (top-left) to MSB (bottom- right). This visualization shows that coarse structural information is concentrated in the high-order MSB planes, while low-order LSB planes consist primarily of noise-like texture. This motivates our non-uniform, utility-aware budget optimization strategy.
  • Bits are Not Created Equal: High-order MSBs contain critical semantic structures, while low-order LSBs resemble noise. Optimizing the budget mathematically ensures that noise is placed where it harms utility the least.

Key Results

  • State-of-the-Art Utility: On face recognition benchmarks (with a privacy budget of $\varepsilon_{total} = 20$), LDP-Slicing achieved 96.68% accuracy on AgeDB-30 and 99.75% on LFW, outperforming existing DP/LDP methods like DCTDP and PEEP while offering a privacy guarantee roughly 4.7$\times$ stricter than block-level alternatives.
  • Robust Attack Resilience: The framework successfully thwarted advanced white-box and black-box reconstruction attacks (including StyleGAN-based inversion), leaving adversaries unable to recover recognizable facial features.
Figure 5. Resilience to a white-box reconstruction attack. The adversary has full knowledge of our pipeline and trains a special- ized two-stage inversion model. From top to bottom εtotal = 5.2, εtotal = 20, εtotal = 58 (a) The original images. (b) original images after LL removed. (c) LDP-Slicing privatized images. (d) LL recovered images. (e) Final recovered images.
  • Zero Overhead: Unlike frequency-domain methods that increase storage requirements by up to 63$\times$, LDP-Slicing outputs standard images with zero storage or transmission overhead ($\times 1$).
  • Edge-Ready Speed: Processing takes just 5.5 ms per image on a consumer-grade Apple M4 chip, making it highly practical for on-device deployment.

Limitations and Future Work

While pixel-level LDP provides robust protection by disrupting spatial correlations, composing the privacy bounds across all pixels mathematically results in a loose worst-case theoretical guarantee for the entire image as a single unit. The authors note that future work will focus on extending this bit-plane slicing paradigm to video streams, where temporal redundancies present both new challenges and opportunities for privacy budget optimization.

Summary updated

Abstract

Local Differential Privacy (LDP) is the gold standard trust model for privacy-preserving machine learning by guaranteeing privacy at the data source. However, its application to image data has long been considered impractical due to the high dimensionality of pixel space. Canonical LDP mechanisms are designed for low-dimensional data, resulting in severe utility degradation when applied to high-dimensional pixel spaces. This paper demonstrates that this utility loss is not inherent to LDP, but from its application to an inappropriate data representation. We introduce LDP-Slicing, a lightweight, training-free framework that resolves this domain mismatch. Our key insight is to decompose pixel values into a sequence of binary bit-planes. This transformation allows us to apply the LDP mechanism directly to the bit-level representation. To further strengthen privacy and preserve utility, we integrate a perceptual obfuscation module that mitigates human-perceivable leakage and an optimization-based privacy budget allocation strategy. This pipeline satisfies rigorous pixel-level $\varepsilon$-LDP while producing images that retain high utility for downstream tasks. Extensive experiments on face recognition and image classification demonstrate that LDP-Slicing outperforms existing DP/LDP baselines under comparable privacy budgets, with negligible computational overhead.
BibTeX
@inproceedings{cao2026ldp,
  author = {Yuanming Cao and Chengqi Li and Wenbo He},
  title = {{LDP-Slicing: Local Differential Privacy for Images via Randomized Bit-Plane Slicing}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2026},
  url = {https://openreview.net/forum?id=6ZfOM52oGR},
}
Figure