FEnc$^2$: Unifying Data Packing for Efficient Private Inference via Convolution and Architecture-Aware Fragment Encoding

2026-06-15 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, quick

Summary

FEnc^2 is a unified, fragment-based encoding framework designed for CKKS-based private convolutional neural network inference, specifically addressing the extreme computational and memory overheads caused by inefficient ciphertext packing in Fully Homomorphic Encryption (FHE). It comprises Conv-aware Encoding, which selects optimal fragment sizes to minimize rotations, and Arch-aware Ct Compression, which restores ciphertext density after feature or channel reduction. This approach optimizes slot utilization, rotation complexity, and ciphertext density, reducing homomorphic operations by one to two orders of magnitude. FEnc^2 achieves significant speedups over the state-of-the-art Orion, with up to 228.83x on GPU and 226.06x on CPU for LeNet on MNIST, and up to 4.55x on GPU and 9.43x on CPU for MobileNet on ImageNet, demonstrating that application-level data layout is a critical architectural design dimension.

Key takeaway

For AI Architects and Machine Learning Engineers designing privacy-preserving inference systems, FEnc^2 offers a critical advancement. Its unified fragment-based encoding framework significantly reduces FHE overheads by optimizing data packing at the application level, complementing existing primitive-level optimizations. You should consider FEnc^2's approach to achieve substantial speedups, potentially up to 228x, and improve memory utilization in your CKKS-based CNN deployments.

Key insights

FEnc^2 unifies data packing for FHE private inference, drastically reducing computational and memory overheads.

Principles

• Inefficient ciphertext packing is a primary FHE overhead.
• Application-level data layout is a first-order architectural design dimension.

Method

FEnc^2 employs Conv-aware Encoding to select optimal fragment sizes and Arch-aware Ct Compression to restore ciphertext density, reshaping encrypted workload structure and reducing homomorphic operations.

In practice

• Optimizes encrypted tensor layout before execution.
• Reduces ciphertext count and workload pressure on hardware.

Topics

• Fully Homomorphic Encryption
• Private Inference
• Convolutional Neural Networks
• Data Packing
• Ciphertext Compression
• CKKS Scheme
• Performance Optimization

Best for: Research Scientist, Computer Vision Engineer, AI Scientist, Machine Learning Engineer, AI Architect

Related on AIssential

• SecureRouter: Encrypted Routing for Efficient Secure Inference — SecureRouter accelerates encrypted Transformer inference via input-adaptive model selection and an MPC-optimized model pool.
• End-to-end encrypted ML inference with Amazon SageMaker AI and FHE — FHE enables ML inference on encrypted data, ensuring privacy from cloud providers like SageMaker AI.
• Sequential KV Cache Compression via Probabilistic Language Tries: Beyond the Per-Vector Shannon Limit — Exploiting the sequential, predictable nature of KV caches significantly surpasses per-vector compression limits.
• Privacy from Symmetry: Orthogonally Equivariant Transformers for LLM Inference — Enforcing architectural symmetry via orthogonal obfuscation and equivariant transformers provides practical privacy for LLM inference.
• Cluster-Level Attention-Guided Parallel Decoding for Masked Diffusion Language Models — CLAD improves MDLMs by committing predictions at a cluster-level, guided by attention, for faster parallel decoding.
• A team of researchers form Meta and Stanford Propose Fast Byte Latent Transformer That Reduces Inference Memory Bandwidth by Over 50% Without Tokenization — New BLT methods drastically cut byte-level model inference memory bandwidth without tokenization or quality loss.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.