Graph Propagated Projection Unlearning: A Unified Framework for Vision and Audio Discriminative Models

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

Summary

Graph-Propagated Projection Unlearning (GPPU) is a new, unified, and scalable algorithm designed for class-level unlearning in deep neural networks, applicable to both vision and audio models. GPPU addresses the need for efficient removal of learned information due to privacy, regulatory compliance, and adaptive system design requirements. The method operates in two phases: first, it uses graph-based propagation to identify class-specific directions in the feature space, and second, it projects representations onto an orthogonal subspace, followed by targeted fine-tuning, to irreversibly remove target class information. Evaluations across six vision datasets and two large-scale audio benchmarks, using architectures like CNNs, Vision Transformers, and Audio Transformers, demonstrate that GPPU achieves 10-20x speedups over prior methods while preserving model utility on retained classes. It consistently reaches near-zero forget-class accuracy post-unlearning and maintains high utility on retained classes.

Key takeaway

For machine learning engineers and researchers building privacy-compliant or adaptable AI systems, GPPU offers a significantly more efficient and effective approach to class-level unlearning. You should consider integrating GPPU for scenarios requiring rapid and precise removal of specific class knowledge, especially in large-scale vision and audio models, as it drastically reduces computational overhead while maintaining high model utility on retained data. This framework enables robust continual unlearning and better adherence to data protection regulations like India's Digital Personal Data Protection Act, 2023.

Key insights

GPPU efficiently unlearns specific classes from deep networks by geometrically projecting features and fine-tuning.

Principles

• Graph propagation refines class boundaries.
• Orthogonal projection removes class-specific features.
• Targeted fine-tuning ensures intrinsic unlearning.

Method

GPPU identifies class-specific "forget directions" via graph-based propagation on features, then projects representations onto the orthogonal subspace and fine-tunes the model's final layers using projection and retention losses.

In practice

• Achieves 10-20x speedup over prior unlearning methods.
• Maintains >93% retention accuracy on non-target classes.
• Applicable to both vision and audio models.

Topics

• Graph-Propagated Projection Unlearning
• Machine Unlearning
• Class-Level Unlearning
• Vision Models
• Audio Models

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

Related on AIssential

• PURGE: Projected Unlearning via Retain-Guided Erasure — PURGE unlearns data by projecting gradients to preserve retain-set performance and erasing hidden representations towards retain distribution.
• Interference-Aware Multi-Task Unlearning — Multi-task unlearning requires mitigating task- and instance-level interference to preserve retained performance and generalization.
• Projected Gradient Unlearning for Text-to-Image Diffusion Models: Defending Against Concept Revival Attacks — PGU hardens diffusion models against concept revival during fine-tuning by projecting gradient updates into a retain-concept-orthogonal subspace.
• Machine Unlearning for the XGBoost Model with Network Intrusion Datasets — XGBoost-Forget enables efficient machine unlearning for tabular network intrusion data, bridging a critical research gap.
• Auditing Machine Unlearning: A Systematic Research on Whether Models Truly Forget — A new auditing framework reliably verifies machine unlearning effectiveness, revealing varying success across different unlearning methods.
• Forgetting That Sticks: Quantization-Permanent Unlearning via Circuit Attribution — Quantization can reverse machine unlearning, necessitating methods that ensure forgetting persists post-compression.

Open in AIssential →

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