Dual Sided Privacy for Machine Learning Inference

140 concepts125 papers2 proposals

The explosion of Machine Learning as a Service (MLaaS) has introduced a fundamental privacy deadlock. When a user queries a cloud-hosted model—whether it is a medical diagnostic tool or a large language model—two competing assets must be protected:

  1. The User's Data Privacy: The input query may contain highly sensitive personal, financial, or medical information that the user cannot risk exposing to the cloud provider.
  2. The Provider's Intellectual Property: The model weights represent massive financial investment and proprietary engineering that the provider cannot distribute directly to the user's edge device.

As highlighted by Liu et al. in Oblivious Neural Network Predictions via MiniONN transformations, standard cloud deployment models force an all-or-nothing compromise. Either the user uploads plaintext data to the server, or the server ships the raw model weights to the client.

To break this deadlock, a rapidly evolving research direction focuses on dual-sided privacy-preserving inference. The goal is to execute complex neural networks in a "double-blind" setting: computing accurate predictions without the server ever seeing the unencrypted input, and without the client ever learning the underlying model weights.


The Core Mechanism: Secret Sharing and MPC

The primary mathematical foundation for dual-sided privacy is Secure Multi-Party Computation (MPC), often implemented via cryptographic Secret Sharing.

To understand why this works, imagine a user has a private input $x$ and a server has a private model weight $w$. Instead of sharing these values directly, they split them into random-looking fragments called "shares." For instance, the user splits $x$ into two numbers, $x_1$ and $x_2$, such that $x_1 + x_2 = x$. The user keeps $x_1$ and sends $x_2$ to the server. Crucially, neither $x_1$ nor $x_2$ reveals anything about the original value of $x$ on its own.

Because of the linear properties of this math, adding two secret-shared values is trivial and requires zero communication between the parties:

$$(x_1 + y_1) + (x_2 + y_2) = (x_1 + x_2) + (y_1 + y_2) = x + y$$

Linear operations—such as the matrix multiplications that dominate the attention layers of Transformers or the convolutional layers of CNNs—can be computed entirely on these isolated shares. The client and server compute their respective parts locally, maintaining perfect confidentiality.


The Non-Linear Bottleneck

If linear operations are cheap and secure, why is private inference notoriously slow? The bottleneck lies entirely in the non-linear activation functions (such as ReLU, GeLU, SiLU, and Softmax).

Cryptographic protocols cannot natively compute non-linearities on secret shares without massive communication overhead. For example, evaluating a standard ReLU function ($f(x) = \max(0, x)$) requires the parties to securely determine whether the hidden value $x$ is greater than zero. This necessitates interactive protocols like Garbled Circuits or Oblivious Transfer, which require the client and server to pass messages back and forth for every single activation node. For a multi-billion parameter model, this communication overhead quickly grinds inference to a halt.


Overcoming the Bottleneck via Polynomial Approximations

To make dual-sided private inference practical, researchers have shifted from trying to compute exact non-linear functions to replacing them with MPC-friendly approximations.

Instead of evaluating a complex transcendental function like Softmax or a sharp piecewise function like ReLU, frameworks substitute them with low-degree polynomials (such as quadratic functions). Polynomials rely exclusively on additions and multiplications, which are significantly faster to compute within MPC frameworks.

Several recent breakthroughs demonstrate the power of this approach:

  • Transformer Inference: In MPCFormer: fast, performant and private Transformer inference with MPC, Li et al. replaced the computationally punishing Softmax and GeLU functions in BERT models with quadratic approximations. By combining these approximations with Knowledge Distillation (KD)—training the approximated model to mimic the outputs of the original plaintext model—the authors recovered virtually all lost accuracy while achieving up to a 5.9x speedup.
  • Large Language Models: Scaling this to generative models, Lu et al. introduced PUMA: Secure Inference of LLaMA-7B in Five Minutes. PUMA designs high-quality mathematical approximations specifically tailored for LLaMA's non-linear layers, enabling secure inference of massive 7-billion-parameter models without requiring expensive retraining.
  • Generative Diffusion Models: The challenge extends beyond text. In CipherDM: Secure Three-Party Inference for Diffusion Model Sampling, Chen et al. optimized the SiLU and Softmax functions essential to image generation, proving that polynomial approximations can make secure multi-party image generation viable.

An alternative take on this philosophy is explored by Wang et al. in Taylor Unswift: Secured Weight Release for Large Language Models via Taylor Expansion. Instead of keeping the model strictly on the server, the authors transform the model weights into Taylor-series parameters. This allows a developer to release a functional model to the public that protects intellectual property and user privacy simultaneously, while intentionally inducing high latency to prevent commercial exploitation.


Alternative Frontiers: Hardware and Information Theory

While cryptographic MPC and Homomorphic Encryption offer mathematical guarantees, other branches of this research direction look to hardware or information theory to bypass the performance penalties of pure cryptography.

Trusted Execution Environments (TEEs)

Instead of encrypting the data mathematically, hardware-based approaches isolate the computation inside a secure enclave on the chip itself. In Towards Confidential and Efficient LLM Inference with Dual Privacy Protection, Yu et al. proposed a hybrid framework that sanitizes sensitive input tokens on a client-side TEE before offloading the heavy lifting to remote cloud GPUs. This minimizes the massive communication bottlenecks typical of pure cryptographic partitioning.

Information-Theoretic Privacy

For highly compressed or quantized models, researchers are exploring frameworks that treat inference as a variant of Private Information Retrieval (PIR). As demonstrated by Deng et al. in Private Inference in Quantized Models, this framing exposes a fundamental mathematical trade-off: there is a fixed privacy budget between model weight confidentiality and user data privacy. Optimizing private inference requires carefully balancing this budget based on the specific security constraints of the application.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 124 papers
  2. Version 1 · 2026-10-04 · 84 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_privacy_preserving_inference_69f3a3e1,
  title = {privacy-preserving inference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/privacy-preserving-inference/rdr_69f3a3e1aae5428d8113ee358b151e73},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_privacy_preserving_inference_69f3a3e1_v2,
  title = {privacy-preserving inference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/privacy-preserving-inference/rdr_69f3a3e1aae5428d8113ee358b151e73/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake