Deep neural networks (DNNs) are built on the arithmetic of real numbers, yet Secure Multiparty Computation (MPC)—the cryptographic framework used to process data without revealing it—operates almost exclusively on integers within large rings or fields. To bridge this gap, researchers use fixed-point arithmetic, where fractional values are scaled into integers. However, every multiplication in this format doubles the scaling factor, requiring a "truncation" step to return the value to its original scale.
In standard computing, truncation is a trivial bit-shift. In MPC, it is a primary bottleneck. Recent research has focused on optimizing this mechanism, moving from expensive deterministic methods to efficient probabilistic truncation, enabling secure inference to scale from simple classifiers to Large Language Models (LLMs).
The Scaling Problem in Secure Arithmetic
When two fixed-point numbers with $f$ bits of fractional precision are multiplied, the result has $2f$ fractional bits. To prevent the bit-width from growing indefinitely and eventually overflowing the cryptographic ring, the parties must divide the result by $2^f$.
As noted by Kanav Gupta and colleagues in the development of SIGMA, performing this truncation "faithfully" (matching cleartext bit-for-bit) requires complex cryptographic circuits. These circuits introduce significant latency and communication overhead, which historically made large-scale deployment like GPT inference impractical.
Probabilistic Truncation and Neural Resilience
The key insight driving recent efficiency gains is that neural networks are inherently stochastic and resilient to small amounts of numerical noise. Instead of computing an exact truncation, researchers like Kyle Bittner and Rafael Dowsley have championed probabilistic truncation.
In this approach, a value $x$ is rounded to either $\lfloor x/2^f \rfloor$ or $\lceil x/2^f \rceil$ based on its proximity to the nearest integer. This statistical rounding is significantly cheaper to compute in MPC. Anders P. K. Dalskov and Marcel Keller demonstrated in Secure Evaluation of Quantized Neural Networks that this method drastically reduces communication overhead in ring-based computation. Because the errors are zero-mean and small, they rarely impact the final classification accuracy of the model.
Bitwise Equivalence for Deep Architectures
While probabilistic methods work for many tasks, deep architectures can suffer from "error drift" if small rounding errors accumulate across hundreds of layers. To solve this, Deevashwer Rathee and the team behind CrypTFlow2 developed protocols that achieve bitwise equivalence to cleartext execution.
By decomposing division into local computations and secure correction terms, they ensured that the secure model behaves exactly like its plaintext counterpart. This precision allowed them to perform secure inference on ImageNet-scale models for the first time, proving that high-capacity models do not necessarily require a trade-off between security and mathematical exactness.
Optimizing the Non-Linear Frontier
Truncation is not just a tool for multiplication; it is a prerequisite for handling non-linear activation functions like ReLU, Sigmoid, and Tanh.
- Stochastic Approximations: Nandan Kumar Jha and Siddharth Garg introduced Circa, which truncates the least significant bits (LSBs) of inputs before they enter a Garbled Circuit. This reduces the number of logic gates needed for ReLU, speeding up inference by 3x.
- Lookup Tables (LUTs): Protocols like Tabula by Max Lam use truncation to downsample high-precision linear outputs into 8-bit indices for secure lookup tables, replacing heavy boolean circuits with efficient memory accesses.
- Scale Management: In recurrent networks, where errors can compound over time steps, Qian Feng and colleagues in OPAF showed that carefully managing numerical scale through intermediate truncation steps is vital for maintaining precision in complex functions like Tanh.
Formalizing Security and Performance
A long-standing debate in the field concerned whether the statistical noise in probabilistic truncation could leak information about the underlying private data. Shubho Sengupta, in the work on Curl, provided formal security proofs for these functionalities. By validating that these efficient arithmetic techniques are secure in the stand-alone model, this research has cleared the path for practitioners to use faster truncation methods without compromising cryptographic rigor.
This evolution—from slow, exact integer arithmetic to fast, "good enough" probabilistic truncation—has been the primary engine allowing MPC to move from simple logistic regression (as seen in iDASH 2019) to the secure execution of multi-billion parameter LLMs.