Current result
Awaiting updateRun the method to compare it with the retrained reference.
Do not trust “deleted.” Measure forgotten.
Machine unlearning aims to remove selected training influence without rebuilding everything. Choose a forget set, compare methods against exact retraining, and watch removal quality collide with retained utility, privacy leakage, and compute cost.
unlearned ≈ retrained without Dᶠverify removal + utility + privacy
Training influence map
Select records to include in the forget set.
Forget confidence
Retain utility
Leakage score
Deletion trajectory
Distance to the exact retrain reference across optimization steps.
Step 0
Verification gate
Not evaluatedA deletion claim needs removal, retained performance, and attack-based evidence.
Retrain distance1.00
Membership AUC0.85
Utility drop0%
Compute1×
Deletion is a comparison, not a ceremony.
The strongest reference trains from scratch without the forget set. Approximate methods are useful only when their outputs, attacks, and retained-task behavior stay acceptably close to that counterfactual.
One score cannot certify forgetting.
Pair forget-set performance with retained utility, membership inference, output distance, and targeted probes. A model can fail one signal while looking clean on another.
Exact deletion is expensive.
Full retraining is the clean baseline, but it repeats almost all original compute.
≈100%
Collateral damage matters.
Destroying broad capability can make forget accuracy look excellent for the wrong reason.
Privacy attacks are evidence.
Membership inference tests whether forgotten examples still look unusually familiar.
Threat models differ.
Record deletion, class removal, poisoning repair, and generative memorization need different probes.
Audit the counterfactual.
Freeze the original checkpoint, forget set, retained set, attack suite, random seeds, and exact retrain baseline before comparing approximate deletion.
Measure the forgotten subset
Inspect loss, confidence, calibration, extraction, and task-specific behavior on the requested records.
Protect unrelated capability
Use held-out retained data and downstream tasks to expose broad degradation disguised as deletion.
Run adversarial verification
Membership inference and targeted extraction can reveal influence that average accuracy misses.
Methods trade guarantees for speed.
Hover or focus each method to expand the mechanism and its verification burden.
Primary sources behind the audit.
The trajectories are illustrative. These papers introduce machine unlearning, SISA training, certified removal, and modern evaluation of approximate unlearning.
FOUNDATIONAL DEFINITION
Machine Unlearning
Frames efficient removal of training influence without full retraining.
SHARDED TRAININGMachine Unlearning with SISA
Introduces sharding, isolation, slicing, and aggregation for efficient deletion.
CERTIFIED REMOVALCertified Data Removal from Machine Learning Models
Develops a statistical indistinguishability guarantee for removal.
EVALUATIONWhen Machine Unlearning Jeopardizes Privacy
Shows why unlearning outputs and privacy attacks require careful joint evaluation.
Export a deletion audit.
Preserve the method, threat model, evidence, and failure threshold. Then see how Superpowers builds inspectable agent tools and caches repeatable computer-use evaluations.
0 of 4 ready