ZK-APEX: Zero-Knowledge Approximate Personalized Unlearning with Executable Proofs

2026-04-16 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Emerging Technologies & Innovation · Depth: Expert, medium

Summary

ZK-APEX is a novel framework for Zero-Knowledge Approximate Personalized Unlearning with EXecutable Proofs, designed to remove the influence of specific data from personalized machine learning models on resource-constrained edge devices. It addresses the challenge of verifying unlearning without exposing private model parameters or data, a critical need for privacy regulations like GDPR. The method combines provider-side sparse masking for targeted data removal with client-side Group-OBS compensation, computed from a block-wise empirical Fisher matrix, to restore personalization accuracy. This curvature-aware update is optimized for low-overhead execution and proof generation using Halo2 ZK-SNARKs. ZK-APEX recovers approximately 99% Top-1 personalization accuracy on Vision Transformer (ViT) classification models and about 70% accuracy on OPT125M generative models. Proof generation for ViT takes approximately 2 hours, which is over $10^{7}\times$ faster than retraining-based verification, with peak memory under 0.7 GB and proof sizes around 400 MB.

Key takeaway

For CTOs and VPs of Engineering deploying personalized AI models on edge devices, ZK-APEX offers a practical solution for regulatory compliance and data privacy. You can implement verifiable unlearning without compromising local model utility or exposing sensitive client data, significantly reducing verification time and computational overhead compared to traditional retraining methods. This framework allows you to confidently meet "right to be forgotten" requests while maintaining model performance and privacy.

Key insights

ZK-APEX enables verifiable personalized machine unlearning on edge devices using ZK-SNARKs and curvature-aware updates.

Principles

• Unlearning requires verifiable compliance without exposing private data.
• Curvature-aware updates can balance forgetting and personalization retention.
• Zero-shot unlearning avoids SGD randomness, enhancing ZKP reliability.

Method

ZK-APEX uses provider-side sparse masking for targeted removal and client-side Group-OBS compensation via a block-wise empirical Fisher matrix. This generates a curvature-aware update, proven correct using Halo2 ZK-SNARKs.

In practice

• Apply sparse masking for targeted data removal.
• Use Group-OBS compensation to restore model performance.
• Leverage ZK-SNARKs for verifiable unlearning on edge devices.

Topics

• Machine Unlearning
• Personalized Unlearning
• Zero-Knowledge SNARKs
• Edge Devices
• Curvature-aware Unlearning

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• Roots Beneath the Cut: Uncovering the Risk of Concept Revival in Pruning-Based Unlearning for Diffusion Models — Pruning-based unlearning in diffusion models is vulnerable to concept revival via side-channel information from pruned weight locations.
• Zero knowledge verification for frontier AI training is possible — A novel zero-knowledge architecture enables verifiable, confidential, and efficient frontier AI training by directly proving GPU floating-point computation.
• Private analytics via zero-trust aggregation — A zero-trust private analytics solution combines one-shot cryptographic aggregation with TEEs for robust, multi-layered data protection.
• From Anchors to Supervision: Memory-Graph Guided Corpus-Free Unlearning for Large Language Models — MAGE enables auditable, corpus-free LLM unlearning using minimal anchors and self-generated supervision.
• Protecting the Undeleted in Machine Unlearning — Perfect retraining in machine unlearning poses significant privacy risks to undeleted data via reconstruction attacks.
• pFedUL: Layer-Aware Federated Unlearning for Personalized Federated Learning — Layer-aware federated unlearning balances data removal completeness with personalization preservation in pFL.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.