🤖 AI

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

This paper introduces Graph-Propagated Projection Unlearning (GPPU), a unified and scalable framework that leverages graph-based propagation and orthogonal projection to efficiently and irreversibly erase class-specific information from both vision and audio discriminative models, achieving 10–20x speedups over existing methods while preserving utility on retained classes.

Original authors: Shreyansh Pathak, Jyotishman Das

Published 2026-04-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shreyansh Pathak, Jyotishman Das

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a super-smart student who has studied for years to recognize thousands of things: apples, oranges, cars, dogs, and even specific voices. This student is so good that they can tell the difference between a Golden Retriever and a Poodle instantly.

But then, a problem arises. Due to privacy laws or a simple request, the student must completely forget how to recognize "Apples." They need to unlearn this specific skill without forgetting how to recognize oranges, cars, or dogs.

In the past, the only way to do this was to send the student back to school and make them re-study everything from scratch, just without the apple pictures. This is slow, expensive, and wasteful. Other methods tried to "brainwash" the student by telling them "Apples are actually Oranges," but this often confused them, making them bad at recognizing Oranges too.

Enter GPPU (Graph-Propagated Projection Unlearning).

Think of GPPU not as a teacher forcing the student to re-study, but as a high-tech librarian and a geometric sculptor working together. Here is how it works, step-by-step:

1. The "Friendship Map" (Graph Propagation)

First, the system looks at all the data the student has learned. Imagine every picture the student has seen is a person at a huge party.

The Problem: Sometimes a picture of an apple is blurry or has a leaf on it, making it look weird.
The GPPU Solution: The system draws a map of who is standing next to whom. It says, "Hey, all the apples are standing in a tight circle near each other." It smooths out the weird, blurry apples by averaging them with their "friends" (other clear apples).
The Result: The system now knows exactly where the "Apple Circle" is located in the room. It identifies the center of that circle. This is the "Forget Direction."

2. The "Magic Wall" (Orthogonal Projection)

Now, the system needs to erase the Apple Circle without destroying the rest of the room.

The Metaphor: Imagine the student's knowledge is a 3D room. The "Apple Circle" is a specific hallway leading to a door labeled "Apple."
The Action: Instead of demolishing the whole building (retraining), GPPU builds an invisible, magical wall perpendicular to that hallway.
The Effect: When the student tries to look at a picture of an apple, their brain hits the wall and bounces off. They can no longer see the "Apple" door. But because the wall is built at a perfect 90-degree angle, it doesn't block the hallway to "Oranges" or "Cars." Those paths remain wide open and clear.

3. The "Fine-Tuning" (Targeted Practice)

Just building the wall isn't enough; the student needs to practice walking around it so it becomes second nature.

GPPU takes the student for a very short, targeted workout (just a few minutes, not months).
It tells the student: "If you see something that looks like an apple, push it away from the Apple Hallway. If you see an orange, keep walking straight."
Because this practice is so specific and short, it takes 10 to 20 times less time than the old methods of retraining the whole student.

Why is this a Big Deal?

It's Fast: It's like using a laser to remove a single stain from a shirt, rather than washing the whole shirt 20 times.
It's Precise: The student forgets apples perfectly (they guess randomly, just like a coin flip) but remains an expert at everything else.
It Works Everywhere: This trick works for Vision (images) and Audio (voices). Whether the student is trying to forget a specific speaker's voice or a specific type of bird song, the "Magic Wall" works the same way.

The Bottom Line

GPPU is a new, super-efficient way to make AI "forget" specific things on demand. It uses geometry and smart mapping to surgically remove unwanted knowledge without damaging the rest of the AI's brain. This is a huge step forward for privacy, allowing us to build AI systems that can adapt and respect user requests without needing to be rebuilt from the ground up every time.

1. Problem Statement

Deep neural networks (DNNs) are increasingly deployed in vision and audio domains, but they face growing regulatory and ethical demands to selectively erase specific learned information (e.g., specific classes or user data) without retraining the entire model.

Challenges: Existing unlearning methods often rely on full retraining (prohibitively expensive) or adversarial optimization (Gradient Ascent), which suffer from poor scalability, high computational costs, and "catastrophic forgetting" (degrading performance on retained classes).
Gap: There is a lack of unified, efficient, and modality-agnostic frameworks that can effectively unlearn class-level information across both visual and audio domains while preserving model utility and ensuring irreversible removal of target data.

2. Methodology: Graph-Propagated Projection Unlearning (GPPU)

GPPU is a two-stage algorithm designed to identify and remove class-specific information geometrically within the feature space.

Stage 1: Graph-Based Forget Direction Identification

Instead of using raw features, GPPU constructs a k-Nearest Neighbor (k-NN) graph over the feature space to smooth intra-class variations and sharpen class boundaries.

Feature Extraction: Deep features are extracted from the pre-trained model.
Graph Construction: A k-NN graph is built based on cosine similarity.
Graph Propagation: A single layer of graph convolution (without learnable weights) is applied to aggregate neighbor features. This denoises the features and refines the manifold structure.
Direction Extraction: For each class $c$ to be forgotten, a forget direction ( $g_c$ ) is computed as the normalized centroid of the smoothed features for that class. This vector represents the principal direction of the class in the feature manifold.

Stage 2: Projection-Based Representation Unlearning

Once the forget directions are identified, GPPU removes the information via orthogonal projection and targeted fine-tuning.

Orthogonal Projection: The model's feature representations are projected onto the subspace orthogonal to the forget directions ( $g_c$ ). This immediately suppresses the ability to classify the target class.
Targeted Fine-Tuning: To ensure the unlearning is intrinsic (not just output masking), the model is fine-tuned on a small subset of parameters (typically the final layers of the feature extractor).
- Projection Loss ( $L_{proj}$ ): Penalizes the magnitude of the projection of features onto the forget directions, forcing features to become orthogonal to $g_c$ .
- Retention Loss ( $L_{retain}$ ): Standard cross-entropy loss on retained classes to preserve utility.
Optimization: The model minimizes a weighted sum of these losses, updating only the final layers to avoid catastrophic forgetting of low-level features.

3. Key Contributions

Unified Geometric Framework: GPPU is the first framework to operate effectively across both Vision (CNNs, ViTs) and Audio (Audio Transformers) modalities using a single geometric approach.
Graph-Based Direction Discovery: It leverages graph propagation to robustly identify class-specific directions, handling noise and intra-class variance better than raw centroid approaches.
High-Fidelity Unlearning: The method achieves near-zero accuracy on forget classes while maintaining high accuracy (>93%) on retained classes.
Exceptional Efficiency: By avoiding full retraining and using lightweight fine-tuning, GPPU achieves 10–20× speedups over state-of-the-art baselines.

4. Experimental Results

The framework was evaluated on six vision datasets (CIFAR-10/100, SVHN, Flowers102, STL-10, FashionMNIST) and two large-scale audio benchmarks (LibriSpeech, VoxCeleb1, SpeechCommands) using architectures like ResNet, ViT, Wav2Vec2, and HuBERT.

Single-Class Unlearning:
- Forgetting: GPPU reduced Forget Accuracy (FA) to ~0.1–0.4% (near random chance), significantly outperforming baselines like Gradient Ascent (GA) and Fisher Forgetting (FF).
- Retention: Retain Accuracy (RA) remained above 90–95%, comparable to the original model.
- Privacy: Achieved minimal Information Leakage (IL) and Membership Inference Attack (MIA) vulnerability.
- Speed: Runtime was 10–20× faster than competing methods (e.g., ~30s vs. ~800s for GA).
Continual Unlearning:
- In scenarios where 10% of classes were sequentially removed, GPPU maintained stability, whereas gradient-based methods suffered from cumulative interference and performance degradation.
Audio Performance:
- GPPU successfully unlearned specific phonemes and speaker identities in audio models, achieving near-perfect forgetting while preserving speech recognition capabilities, outperforming specialized audio unlearning methods like Quantum-Inspired Audio Unlearning (QIAU).
Representation Analysis:
- t-SNE Visualizations: Showed the forgotten class cluster dispersing and overlapping with neighbors, while retained classes remained distinct.
- Linear Probing: A fresh linear classifier trained on the unlearned features achieved near-random accuracy (~10% on CIFAR-10), proving that the discriminative subspace was irreversibly destroyed, not just masked.

5. Significance

Scalability: GPPU addresses the computational bottleneck of machine unlearning, making it feasible for large-scale, real-world applications.
Modality Agnostic: By relying on geometric properties of feature manifolds rather than modality-specific heuristics, it provides a unified solution for diverse deep learning tasks.
Regulatory Compliance: The method offers a principled, efficient way to comply with "Right to be Forgotten" regulations (GDPR) without compromising the utility of the remaining model.
Future Directions: The framework sets a foundation for continual unlearning, federated learning settings, and cross-modal unlearning applications.

In summary, GPPU represents a significant advancement in machine unlearning by combining graph-based manifold reasoning with efficient projection-based optimization, delivering a solution that is simultaneously fast, effective, and robust across vision and audio domains.