GeoForget

GeoForget is a mathematically rigorous framework for machine unlearning, leveraging Riemannian geometry to overcome the limitations of standard Euclidean optimization. This repository provides code and methodology to safely and effectively erase information from deep learning models while maximizing utility.

---

🚀 Core Idea: "Rails vs. Anchors"

• Standard unlearning methods (like L2UL or Finetuning) treat all parameters equally (Euclidean distance), often damaging model utility by "dragging" parameters off the manifold of valid solutions.

• GeoForget redefines the problem using the Fisher-Rao Metric:

  • Rails: Instead of freezing weights (EWC "anchors"), the Fisher Information Matrix builds "rails" guiding optimization, so the model slides along the curvature of the "Remain Data" manifold.
  • Engine: A targeted adversarial attack pushes the model away from the forget class, while a KL-divergence term safely bounds the unlearning energy.

---

🌟 Features

• Riemannian Optimization: Uses the Fisher Information Matrix (FIM) to measure geodesic (true curved) distance in parameter space.
• Guided Surgery: Automatically identifies and locks "innocent" parameters (high Fisher info) while releasing "guilty" ones (low Fisher info) for modification.
• Safety Bounded: Replaces unbounded Cross-Entropy with KL-Divergence to prevent gradient explosion during unlearning.
• Adversarial Robustness: Incorporates L2-bounded adversarial training to ensure the forgotten class is truly erased.

---

📐 Mathematical Formulation

In standard deep learning, parameters are treated as points in a flat Euclidean space (all changes are equal).

GeoForget treats parameters as points on a curved statistical manifold, equipped with the Fisher Information Matrix (FIM) as the Riemannian metric.

• Approximated Geodesic Distance:

  (θ - θ*)ᵗ · Fisher · (θ - θ*)
  

• Optimization Objective: Minimize:

  Total Loss = Loss_Forget + λ_adv * Loss_Adv + λ_geo * Geodesic_Dist
  

  • Forgetting Loss: Push forget class distribution toward uniform noise (KL Divergence).
  • Adversarial Loss: Robustly pushes decision boundary away.
  • Geodesic Pull: Restores and locks "Remain" knowledge via Fisher Information.

---

🛠️ Usage

Installation:
Clone the repository and install dependencies (standard PyTorch stack).

Run GeoForget Unlearning Example:

python main.py \
  --method geoforget \
  --unlearn_rate 1e-4 \
  --adv_lambda 0.1 \
  --geo_lambda 10.0 \
  --description "fisher_rao_preservation"

Parameters

• --geo_lambda: Controls Riemannian constraint strength. Higher = stricter preservation of remain accuracy.
• --adv_lambda: Controls "push" force away from forget class.
• --adv_eps: Radius of adversarial sphere (default: 0.4).

---

📊 GeoForget Results

Dataset	Remain Acc	Forget Acc	MIA Privacy Score
CIFAR-10	92.8%	0.0%	0.87
CIFAR-100	73.1%	0.5%	0.71
VGG Face	89.2%	0.0%	0.78

Note: "MIA Privacy Score" is normalized (1.0 = Indistinguishable from Retrained Model). Higher is better.

GeoForget demonstrates superior privacy-utility trade-offs on complex manifolds like Faces and CIFAR-100.

License

Distributed under the MIT License. See LICENSE for more information.

collaborators: Mohamed Ansaff Abdul Salam, Mohamed Nawran