PRISM: Sensitivity-Aware PolynoMial PRuning for EffIcient Neural Network Encryption

2026-07-20 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, AI Hardware Optimization · Depth: Expert, quick

Summary

PRISM introduces Polynomial-Sensitivity-Aware Pruning (PSAP), a structured pruning method designed for efficient and reliable neural network inference under homomorphic encryption (HE). PSAP scores filters based on weight magnitude, polynomial activation sensitivity, and rotation cost, concentrating pruning in fault-tolerant regions. This approach significantly enhances reliability, limiting catastrophic accuracy drops (>10 pp) to at most two layers in PSAP-pruned models, compared to 5-14 layers for magnitude-pruned baselines. This reduces worst-case vulnerability by up to 29 times under int32 bit-flip injection. Direct CKKS encrypted fault injection indicates a safe operating boundary near BER~ 10^{-5}. The method also improves efficiency, reducing Halevi--Shoup rotations by up to 45.2% on ResNet-32 and lowering multiplicative depth from 66 to 56 levels, enabling leveled inference without bootstrapping. Fault-critical structural layers account for only 1.1% of parameters, allowing targeted hardening.

Key takeaway

For AI Security Engineers deploying neural networks with homomorphic encryption, you should integrate Polynomial-Sensitivity-Aware Pruning (PSAP) to significantly enhance model reliability against bit-flip errors. This method reduces worst-case vulnerability by up to 29 times and allows for selective hardening of only 1.1% of parameters, minimizing overhead. Consider PSAP to achieve a safe operating boundary near BER~ 10^{-5} while also gaining efficiency through reduced rotations and multiplicative depth.

Key insights

PSAP enhances homomorphic encrypted neural network reliability and efficiency by pruning fault-tolerant regions based on sensitivity and rotation cost.

Principles

• Pruning can be reliability-aware.
• Joint scoring improves fault tolerance.
• Selective hardening minimizes overhead.

Method

PSAP scores neural network filters by weight magnitude, polynomial activation sensitivity, and rotation cost to concentrate pruning in fault-tolerant regions, reducing vulnerability in homomorphic encrypted inference.

In practice

• Apply PSAP for HE model deployment.
• Target 1.1% critical layers for hardening.
• Reduce rotations for ResNet-32 HE.

Topics

• Homomorphic Encryption
• Neural Network Pruning
• Model Reliability
• CKKS Encryption
• Bit-Flip Errors
• PSAP Algorithm

Best for: Research Scientist, AI Scientist, AI Security Engineer, AI Hardware Engineer

Related on AIssential

• TGHE: Template-based Graph Homomorphic Encryption for Privacy-Preserving GNN Inference in Edge-Cloud Systems — Ego-centric graph processing with template-based packing enables efficient, privacy-preserving GNN inference on large dynamic graphs.
• Mix-and-Match Pruning: Globally Guided Layer-Wise Sparsification of DNNs — Coordinating existing pruning signals through architecture-aware, layer-wise sparsification yields more efficient and reliable DNN compression.
• Physically Consistent Null Space Alignment for Detection of Low-Magnitude False Data Injection Attacks — PCNSA detects stealthy FDIAs by preserving geometric alignment between physical and measurement-derived pseudo-null spaces through specialized preprocessing.
• Ciphertext- and Polynomial-Level Optimization for Fully Homomorphic Encryption — Multi-level FHE compilation, optimizing at both ciphertext and polynomial granularities, significantly boosts performance and reduces memory.
• PEAR: Supervised Fine-Tuning Optimized for RL — Optimizing SFT for RL requires aligning offline and online learning, while layer pruning severely impacts LLM generative reasoning.
• Unified Static-Dynamic Pruning for Efficient LLM Inference — SPDP unifies static and dynamic pruning with co-designed GPU kernels and data formats for efficient LLM inference.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.