Jaguar: Fast Private Neural Network Inference with Power-of-Two Homomorphic Arithmetic
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.
Nadia: I'm Nadia.
Elias: And I'm Elias, here with Priya.
Nadia: We're looking at the paper: "Jaguar: Fast Private Neural Network Inference with Power-of-Two Homomorphic Arithmetic".
Elias: Here's the gist of the paper. Hybrid HE/2PC private CNN inference remains bottlenecked by prime-modulus homomorphic arithmetic in convolution and by a precision flow that runs ReLU at doubled bitwidth before invoking a separate truncation protocol Jaguar, a system built on a single design choice—a power-of-two ciphertext ring—that addresses both.
Nadia: Addressing Bottlenecks in Inference The paper introduces Jaguar to resolve the bottlenecks in hybrid HE/2PC private CNN inference, which are identified as prime-modulus homomorphic arithmetic in convolution and a precision flow issue involving ReLU activation.
Elias: The authors propose a power-of-two ciphertext ring as the central design choice to enable SPA-CONV and exact ciphertext-side truncation.
Nadia: This choice allows for coefficient-domain convolution evaluation through scalar–polynomial accumulation instead of NTT or polynomial multiplication.
Title and authors: Nadia: So we're looking at this paper, "Jaguar: Fast Private Neural Network Inference with Power-of-Two Homomorphic Arithmetic," and the authors are Yewon Jeong, Nayoung Jung, Hyeri Roh, and Woo-Seok Choi from Seoul National University. Elias, what's the main takeaway here about what they're trying to fix?
Elias: Well, essentially this paper addresses two major bottlenecks in hybrid homomorphic encryption and two-party computation for private neural network inference. They say the problem comes down to using prime-modulus arithmetic when you need convolution, and also a precision flow issue involving the ReLU activation function.
Priya: That sounds like a lot of moving parts, but for us in privacy research, I'm more interested in how they define those bottlenecks. Are we talking about something that makes the computation too slow or just too expensive on top of the privacy loss?
Nadia: Exactly, Priya. It's about making inference faster and more practical without sacrificing the security guarantees of using hybrid encryption and two-party computation. They are proposing a specific design choice to solve both those problems.
Elias: That central design choice is using a power-of-two ciphertext ring, which they argue fixes the issue with prime moduli for convolution arithmetic and also handles that precision flow problem by allowing exact truncation on the ciphertext side.
Priya: So, they're suggesting that sticking to a power-of-two structure is the key to unlocking more efficient private inference for complex models like CNNs?
The paper's summary: Nadia: To break this down further, the core idea of "Jaguar: Fast Private Neural Network Inference with Power-of-Two Homomorphic Arithmetic" is replacing the standard approach with a power-of-two ciphertext ring. This single choice allows them to implement something called SPA-CONV, which is a coefficient domain convolution kernel.
Elias: SPA-CONV means they evaluate convolutions as sums of scalar polynomial products and coefficient shifts instead of doing the traditional NTT or polynomial multiplication that usually happens in those prime-modulus setups. It's a complete overhaul of how they handle the convolution layer.
Priya: So, instead of relying on complex polynomial math for every single convolution step, they are using something much simpler to compute over the ciphertext structure? What does that actually mean for the data we’re processing?
Nadia: It means you get an exact way to do things in the coefficient domain. They say this lets them perform exact ciphertext-side truncation using just local right shifts, which means they don't need that extra post-ReLU truncation protocol anymore.
Elias: That's a big deal because it’s zero communication for that part of the process, and it lets you run the ReLU function at the target fixed-point precision directly without having to double up on the bitwidth first.
Priya: So they are eliminating two separate, potentially wasteful steps—the convolution calculation and that extra truncation step—by choosing this power-of-two ring structure?
The paper's improvements: Nadia: Let's talk about what they actually improved, because the results on ImageNet models are pretty compelling. The main improvements center around making the inference flow much smoother and faster. They introduced a specific flow for convolution, truncation, and decryption that works better than before.
Elias: Specifically, they focus on how the power-of-two arithmetic substrate handles things. They adopt a choice where q equals two Q and p equals two P, which ensures the BFV scale is also a power of two. This single choice satisfies several requirements at once, including Theorem one by construction and aligning the homomorphic outputs with the Z2 l share domain for subsequent two-party computation operations <ref:2606.11827#pg2>.
Priya: That aligns with what I was hearing about their truncation method earlier—it seems they are using bit masking instead of modular reduction, which is a key difference in how you handle noise and precision. What does that mean for the actual accuracy of the results?
Nadia: It means they aren't just speeding things up; they are achieving exactness where it matters. They say this choice replaces modular reduction with bit masking, which is a fundamental shift in the arithmetic backend.
Elias: And to keep things efficient during decryption, they use NTT-assisted decryption. They recover the NTT exactly for that single multiplication step during decryption by using an auxiliary NTT prime, but they do it without messing up the power-of-two protocol substrate itself.
Priya: So, the claim is that this entire setup—the SPA-CONV kernel and the exact truncation—leads to significant performance gains across ResNet models like ResNet-eighteen ResNet-fifty and MobileNetV2 <ref:2606.11827#pg1,ResNet-18, ResNet-50, and MobileNetV2>.
Conclusion: Nadia: So we've covered how they use this power-of-two arithmetic substrate to get SPA-CONV for convolution and exact ciphertext truncation. The paper shows that their system achieves substantial speedups over existing methods like Cheetah and Rhombus under both LAN and WAN conditions.
Elias: The performance metrics are pretty telling, showing a two point zero seven to three point seven two times lower end-to-end latency compared to Cheetah, and twenty point one six to three point thirty six times lower than Rhombus. These gains are most pronounced on ResNet-fifty with a reduction of three point seven two times on LAN and two point seven seven times on WAN <ref:2606.11827#pg1>.
Priya: From a privacy standpoint, the benefit isn't just speed; it's that they’ve eliminated the need for that separate post-ReLU truncation protocol entirely, which is something I think makes sense because it simplifies the overall protocol flow while maintaining precision.
Nadia: It really does. The system also handles output sharing efficiently through grouped export, packing coefficients into a shared record consisting of one shared symbol and multiple values for b, reducing communication cost proportional to G N plus E instead of E N.
Elias: And they keep the decryption complexity at O(N log N) by using that auxiliary NTT prime locally during the decryption step while preserving the main protocol structure.
Priya: So, in summary, "Jaguar: Fast Private Neural Network Inference with Power-of-Two Homomorphic Arithmetic" presents a concrete way to reduce latency and communication overhead in hybrid HE/2PC private CNN inference by fundamentally changing the arithmetic substrate to power-of-two <ref:2606.11827#pg1,Inference with Power-of-Two Homomorphic Arithmetic>.
Nadia: That's the gist of it. It’s a specific engineering fix that tackles both convolution bottlenecks and precision issues in one go.
Yewon Jeong, Nayoung Jung, Hyeri Roh, Woo-Seok Choi
Seoul National University
cs.CR
Submitted: 2026-06-10
Updated: 2026-11-03
Comments: Accepted at NeurIPS 2026. Camera-ready version
Code: https://github.com/Alibaba-Gemini-Lab/OpenCheetah
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 94/100
The gist: Hybrid HE/2PC private CNN inference remains bottlenecked by prime-modulus homomorphic arithmetic in convolution and by a precision flow that runs ReLU at doubled bitwidth before invoking a separate…
Key concepts
- Power-of-Two Arithmetic Substrate
- This is a specific choice where the ciphertext ring parameters are powers of two (q=2Q, p=2P). This single choice simplifies the entire system by replacing complex modular reduction with simple bit masking and natively supporting scalar–polynomial products, which is crucial for efficient convolution.
- SPA-CONV
- This is a new convolution kernel that evaluates dense and depthwise convolutions as sums of shifted scalar–polynomial products. It avoids polynomial multiplication (NTT) entirely in the linear path, making the convolution step much faster and more efficient within the power-of-two arithmetic framework.
- Exact Pre-ReLU Truncation
- This technique performs truncation directly on the ciphertext before applying ReLU, eliminating a separate post-ReLU protocol. It uses local right shifts to run ReLU at the target fixed-point precision, which saves computation and communication compared to previous methods that required a separate truncation step after activation.
Terminology
Summary
Hybrid HE/2PC private CNN inference remains bottlenecked by prime-modulus homomorphic arithmetic in convolution and by a precision flow that runs ReLU at doubled bitwidth before invoking a separate truncation protocol Jaguar, a system built on a single design choice—a power-of-two ciphertext ring—that addresses both. p. 0 Abstract
Hybrid HE/2PC private CNN inference remains bottlenecked by prime-modulus
homomorphic arithmetic in convolution and by a precision flow that runs ReLU
at doubled bitwidth before invoking a separate truncation protocol. We present
Jaguar, a system built on a single design choice—a power-of-two ciphertext
ring—that addresses both.Where the paper says it
Addressing Bottlenecks in Inference
The paper introduces Jaguar to resolve the bottlenecks in hybrid HE/2PC private CNN inference, which are identified as prime-modulus homomorphic arithmetic in convolution and a precision flow issue involving ReLU activation. The authors propose a power-of-two ciphertext ring as the central design choice to enable SPA-CONV and exact ciphertext-side truncation. p. 0 We present
Jaguar, a system built on a single design choice—a power-of-two ciphertext
ring—that addresses both. The choice enables SPA-CONV, a coefficient-domain
convolution kernel that replaces NTT-centric polynomial multiplication with scalar–
polynomial accumulation, and an exact ciphertext-side truncation by local right
shifts that lets ReLU run directly at the target fixed-point precision and eliminates
the post-ReLU truncation protocol. We present
Jaguar, a system built on a single design choice—a power-of-two ciphertext
ring—that addresses both. The choice enables SPA-CONV, a coefficient-domain
convolution kernel that replaces NTT-centric polynomial multiplication with scalar–
polynomial accumulation, and an exact ciphertext-side truncation by local right
shifts that lets ReLU run directly at the target fixed-point precision and eliminates
the post-ReLU truncation protocol.Where the paper says it
Where the paper says it
Core Technical Contributions
Jaguar's technical contributions center around three main innovations. These include:
A power-of-two HE arithmetic substrate for hybrid HE/2PC private CNN inference
SPA-CONV, a coefficient-domain convolution kernel that evaluates pointwise, dense, and depthwise CNN convolutions as sums of scalar–polynomial products and coefficient shifts
Exact pre-ReLU ciphertext-side truncation, a zero-communication operation that removes the separate post-ReLU truncation protocol used by all prior hybrid HE/2PC systems
Mechanism for Power-of-Two Arithmetic
The system adopts the choice of q = 2Q and p = 2P, which ensures that the BFV scale ∆ is also a power of two. This single choice satisfies three requirements: it satisfies Theorem 1 by construction, aligns HE outputs with the Z2l share domain for subsequent 2PC operations, and replaces modular reduction with bit masking. p. 3 Jaguar adopts q = 2Q, p = 2P , so the BFV scale ∆ = q/p = 2Q−P is also a
power of two. This single choice does three things: 1) Satisfies Theorem 1 by construction; A f-bit
ciphertext truncation is exact iff 2
f
|∆. Jaguar cannot use the standard NTT-friendly prime backend for convolution. The convolution kernel must be expressed in terms the power-of-two ring supports natively—
scalar–polynomial products, coefficient shifts, additions, and bit operations—which Sec.Where the paper says it
Where the paper says it
Performance Gains and Protocol Flow
The proposed protocol involves a specific flow for convolution, truncation, and decryption. The SPA-CONV kernel evaluates convolution as sums of shifted scalar–polynomial products without requiring plaintext–ciphertext polynomial multiplication or NTT in the linear path. p. 1 Prior coefficientencoded systems used a power-of-two plaintext/share ring but kept the ciphertext backend
on NTT-friendly primes. • SPA-CONV, a coefficient-domain convolution kernel that evaluates pointwise, dense, and
depthwise CNN convolutions as sums of scalar–polynomial products and coefficient shifts,
with no plaintext–ciphertext polynomial multiplication, NTT, or modular prime reduction. • SPA-CONV, a coefficient-domain convolution kernel that evaluates pointwise, dense, and
depthwise CNN convolutions as sums of scalar–polynomial products and coefficient shifts,
with no plaintext–ciphertext polynomial multiplication, NTT, or modular prime reduction. • Exact pre-ReLU ciphertext-side truncation, a zero-communication operation that removes
the separate post-ReLU truncation protocol used by all prior hybrid HE/2PC systems and
reduces the bitwidth at which ReLU is evaluated.Where the paper says it
Where the paper says it
Decryption Efficiency
To maintain efficiency while avoiding NTT in the linear path, Jaguar employs NTT-assisted decryption. This client-local technique recovers NTT exactly where it is useful—during the single cˆ′1·sk multiplication during decryption—via an auxiliary NTT prime without disturbing the power-of-two protocol substrate. p. 2 • NTT-assisted decryption, a client-local technique that recovers NTT exactly where it is
genuinely useful—the single cˆ
′
1
· sk multiplication during decryption—via an auxiliary NTT
prime, without disturbing the power-of-two protocol substrate. 2 Preliminaries
Notations. The choice enables SPA-CONV, a coefficient-domain
convolution kernel that replaces NTT-centric polynomial multiplication with scalar–
polynomial accumulation, and an exact ciphertext-side truncation by local right
shifts that lets ReLU run directly at the target fixed-point precision and eliminates
the post-ReLU truncation protocol. Where NTT remains genuinely useful—at the
client, for the single polynomial multiplication during decryption—we recover it
through an auxiliary NTT prime, preserving the power-of-two protocol substrate
while keeping decryption O(N log N).Where the paper says it
Where the paper says it
Experimental Results and Comparison
Experiments on ImageNet-scale models show significant performance improvements over prior systems. Jaguar achieves 2. p. 0 Where NTT remains genuinely useful—at the
client, for the single polynomial multiplication during decryption—we recover it
through an auxiliary NTT prime, preserving the power-of-two protocol substrate
while keeping decryption O(N log N). On ImageNet-scale ResNet-18, ResNet-50,
and MobileNetV2 with AVX disabled, Jaguar achieves 2. On ImageNet-scale ResNet-18, ResNet-50,
and MobileNetV2 with AVX disabled, Jaguar achieves 2. 72× lower end-toend latency than Cheetah and 2. 72× lower
end-to-end latency than Cheetah and 2. 36× lower than Rhombus under both LAN and WAN
settings. Rhombus is specifically optimized for the pointwise-convolution and MVM components that dominate ResNet-50’s communication; its integration replaces these in Cheetah but leaves
dense 3×3 convolution and the nonlinear stage untouched. This is why Rhombus’s communication
advantage on ResNet-50 is largest in absolute terms but does not translate proportionally into
latency: Rhombus is 1. The bottleneck is no longer 2PC—it is HE convolution. On ResNet-50 under LAN, our profiling
of Cheetah [16, 23] shows convolution alone takes 96. Contributions. We implement Jaguar on ImageNet-scale ResNet-18, ResNet-50, and MobileNetV2
under both LAN and WAN. The nonlinear
side adds reduced ReLU input bitwidth and eliminated post-ReLU truncation, improving both
compute-dominated LAN and communication-sensitive WAN latency. Jaguar’s power-of-two arithmetic is well-suited to AVX-512 [41, 42], since modular
reduction and ciphertext-side truncation map to vectorized bit masking and arithmetic shifts.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Kernel Complexity Analysis
The complexity analysis of the SPA-Conv kernel shows that dense 3 × 3 convolution scales as Θ(NBCoutCinK2), pointwise convolution scales as Θ(NBCoutCin), and depthwise convolution scales as Θ(NBCoutK2). This complexity analysis separates the kernel from the boundary conversion cost, which is analyzed separately in Section K. p. 4 The output crosses the HE→2PC boundary already at target precision (Section 4). The kernel covers
all CNN convolution types: pointwise sets ΩK = {(0, 0)} (channel-wise scalar accumulation with no
spatial shifts); dense 3 × 3 accumulates nine shifted neighborhoods per input channel; depthwise uses
the same shifted accumulation pattern but with C(o) = {o}. Gazelle [13] uses SIMD with a prime plaintext modulus; Cheetah [16]
introduces coefficient-encoded, rotation-free linear protocols with lean OT-based nonlinearities;
CrypTFlow2 [15] focuses on faithful 2PC primitives; Rhombus [26] reduces output communication for matrix-vector multiplication via input-output packing; Falcon [32] specializes packing for
depthwise convolutions. Orthogonally, PrivCirNet [33] modifies the model to reduce HE linear cost.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Power-of-two arithmetic substrate for hybrid HE/2PC private CNN inference allows for SPA-CONV, which evaluates convolution as
sums of shifted scalar–polynomial products without ever multiplying two polynomials, NTT, or modular prime reduction.
This enables the system to performexact ciphertext-side truncation by local right shifts that lets ReLU run directly at the target fixed-point precision and eliminates the post-ReLU truncation protocol.
-
Exact pre-ReLU ciphertext-side truncation reduces communication significantly, as
Jaguar adopts Cheetah’s LWE export compression: LWE ciphertexts extracted from the same RLWE ciphertext share their component, transmitted only once,
leading to a reduction in communication bits compared to prior systems. -
SPA-CONV provides optimized kernel performance where
Dense 3 x 3 convolution scales as Theta(NBCoutCinK2), pointwise convolution scales as Theta(NBCoutCin), and depthwise convolution scales as Theta(NBCoutK2),
allowing for faster inference on ImageNet-scale models by replacing NTT-centric polynomial multiplication withscalar–polynomial accumulation.
-
NTT-assisted decryption recovers asymptotic efficiency without affecting the protocol substrate, using an auxiliary NTT prime where
the single c'1 · sk multiplication during decryption is handled locally
to keep the client reconstruction time at O(N log N). -
The system achieves superior end-to-end latency and communication compared to existing baselines, demonstrating
2.07–3.72× lower end-toend latency than Cheetah and 2.16–3.36× lower than Rhombus
on ResNet-18, ResNet-50, and MobileNetV2 under both LAN and WAN settings with AVX disabled. -
The protocol allows for efficient handling of output sharing via grouped export, where
Jaguar packs them into a grouped export record consisting of one shared aˆ, multiple b values,
reducing communication proportional toGN + E
instead of the EN required by independent exporting.
Abstract
Hybrid HE/2PC private neural network inference remains bottlenecked by prime-modulus homomorphic arithmetic in convolution and by a precision flow that runs ReLU at doubled bitwidth before invoking a separate truncation protocol. We present Jaguar, a system built on a single design choice--a power-of-two ciphertext ring--that addresses both. The choice enables SPA-Conv, a coefficient-domain convolution kernel that replaces NTT-centric polynomial multiplication with Scalar-Polynomial Accumulation, and an exact ciphertext-side truncation by local right shifts that lets ReLU run directly at the target fixed-point precision and eliminates the post-ReLU truncation protocol. Where NTT remains genuinely useful--at the client, for the single polynomial multiplication during decryption--we recover it through an auxiliary NTT prime, preserving the power-of-two protocol substrate while keeping decryption O(N log N). On ImageNet-scale ResNet-18, ResNet-50, and MobileNetV2 with AVX disabled, Jaguar achieves 2.07-3.72x lower end-to-end latency than Cheetah and 2.16-3.36x lower than Rhombus, with 1.16-1.76x lower communication than Cheetah.
Sources
- Flash: A Hybrid Private Inference Protocol for Deep CNNs with High Accuracy and Low Latency on CPU
- Impala: Low-Latency, Communication-Efficient Private Deep Learning Inference
Related papers
- From Strategy Search to Policy Learning: MCTS-Guided Adversarial Co-Evolution for LLM Safety Alignment
- Privacy Leakage from Recurrent State in LLM Inference
- Incrementally Verifiable Oracle Aided Computation
- Privilege and confidentiality in generative AI workflows
- BadMoE: Backdooring Mixture-of-Experts LLMs via Routing-Aware Triggers and Task-Underutilized Experts
- CoVault: Secure, Scalable Analytics of Personal Data