Research corpus
27 papers, 2015–2026. Every bibliographic record verified against a primary source; every substantive claim traced to a named section that was actually read.
Reading paths
Arguments no single paper states, assembled from several that each state part of it. Start here rather than with the paper list.
-
The collapse of the second-order programme
6 papers
Influence functions and Newton corrections dominate the theory of unlearning. Why does none of that machinery reach a ResNet, and at exactly which step does it break?
-
Why matching the retrained model is a convention, not a definition
5 papers
Almost every unlearning paper measures success as similarity to a model retrained without the forget set. Three independent lines argue that this is not enough. What survives?
-
Why SCRUB's privacy numbers are weaker evidence than they look
5 papers
SCRUB reports basic-MIA accuracy near 50% and reads it as a successful defence. Three orthogonal problems say that number carries less weight than it appears to. What may we actually claim after reproducing it?
-
Unlearning was never only about privacy
4 papers
The modern literature opens with the right to be forgotten. The founding paper does not. What was lost, and what recovered it?
-
One method, three papers: relabeling
3 papers
A short lineage worth following end to end, because it shows how a simple idea accretes machinery -- and because implementing the first gives us most of the third.
Papers
-
Towards Making Systems Forget with Machine Unlearning
Coined "machine unlearning". Converts a learning algorithm into a SUMMATION FORM -- a small number of sums of efficiently computable transformations of the training samples, on which the model depends only through those…
-
Making AI Forget You: Data Deletion in Machine Learning
Defines data deletion as a distributional condition: a deletion operation R_A(D, A(D), i) is valid if A(D_{-i}) and R_A(D, A(D), i) are EQUAL IN DISTRIBUTION. Observes that every algorithm supports deletion trivially by…
-
Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networks
Defines forgetting as making the distribution of scrubbed weights indistinguishable from that of weights trained without the forget set, formalised by the Forgetting Lagrangian. Derives an optimal scrubbing function und…
-
Forgetting Outside the Box: Scrubbing Deep Networks of Information Accessible from Input-Output Observations
Linearises the network around its pretrained weights in the Neural Tangent Kernel regime, then derives a closed-form "one-shot forgetting" vector that jumps from the weights obtained by training on D to those that would…
-
Certified Data Removal from Machine Learning Models
Defines certified removal by bounding the max-divergence between the removal mechanism's output and retraining without the point. Achieves it for regularised linear models by a one-step Newton update on the removed poin…
-
Machine Unlearning
Exact unlearning by construction, achieved at TRAINING time rather than post hoc. Partition the data into S disjoint shards, train one constituent model per shard in complete isolation (no gradient sharing), further spl…
-
When Machine Unlearning Jeopardizes Privacy
Unlearning produces TWO models, the original and the unlearned, and the discrepancy between them is itself a leak. Proposes a membership inference attack that queries both, concatenates or otherwise aggregates the two p…
-
Amnesiac Machine Learning
Proposes THREE methods. "Naive Retraining" continues training on D \ S relying on catastrophic interference. "Unlearning" relabels the sensitive data with random incorrect labels and retrains for a few iterations. "Amne…
-
Approximate Data Deletion from Machine Learning Models
The Projective Residual Update for linear (and via IRLS, logistic) regression. Uses a classical leave-one-out residual identity and the hat matrix to compute what the DELETED model would predict on the removed points wi…
-
Descent-to-Delete: Gradient-Based Methods for Machine Unlearning
Gradient-based unlearning for ERM under an ongoing SEQUENCE of add/delete updates rather than a single deletion. Introduces a publishing function separating the unlearner's secret output from the publicly released model…
-
Remember What You Want to Forget: Algorithms for Machine Unlearning
Introduces DELETION CAPACITY -- the maximum number of samples that can be unlearnt while keeping excess POPULATION risk below a threshold -- and shows unlearning is strictly separated from differential privacy. The unle…
-
Membership Inference Attacks From First Principles
Two contributions. First, a methodological argument: membership inference attacks should be evaluated by true-positive rate at LOW false-positive rates and plotted on log-scale ROC, not by average-case accuracy or AUC,…
-
Towards Adversarial Evaluations for Inexact Machine Unlearning
Argues existing unlearning evaluations are inadequate: matching a particular retrained model is neither necessary (a randomly initialised net has no forget-set information yet is far from the retrained model) nor suffic…
-
On the Necessity of Auditable Algorithmic Definitions for Machine Unlearning
Introduces FORGING: a map that rewrites a training log so the same parameter trajectory is explained by a different dataset, one not containing the point to be forgotten. If such a map exists, a model owner can claim un…
-
Unrolling SGD: Understanding Factors Influencing Machine Unlearning
Proposes VERIFICATION ERROR, the L2 distance between an approximately unlearned model and a naively retrained one, arguing it subsumes a large class of unlearning metrics. Unrolls SGD by Taylor series to derive SINGLE G…
-
Can Bad Teaching Induce Forgetting? Unlearning in Deep Networks Using an Incompetent Teacher
Two teachers and one student. The student is initialised from the fully trained model; the competent teacher IS that model; the incompetent teacher is a randomly initialised network. A per-example hard switch pulls the…
-
Model Sparsity Can Simplify Machine Unlearning
Two contributions. "Prune first, then unlearn": sparsify the trained model (one-shot magnitude pruning, OMP, by default) then apply an existing unlearning method, motivated by a bound showing gradient-ascent unlearning…
-
Towards Unbounded Machine Unlearning
Casts unlearning as teacher-student distillation where the student selectively DISOBEYS the original model (the teacher) on the forget set while obeying it on the retain set. Departs from the approximate-unlearning trad…
-
SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation
Compute a gradient-based WEIGHT SALIENCY MAP by hard-thresholding the magnitude of the forget-set loss gradient at the original weights, then restrict any unlearning update to the salient weights while freezing the rest…
-
Fast Machine Unlearning without Retraining through Selective Synaptic Dampening
Identify parameters disproportionately important to the forget set using the diagonal Fisher information, then DAMPEN rather than prune them, by a factor proportional to how specialised they are. No retraining, no addit…
-
Inexact Unlearning Needs More Careful Evaluations to Avoid a False Sense of Privacy
Categorises unlearning MIAs (U-MIAs) into POPULATION U-MIAs, where one attacker is trained on outputs pooled across different examples, and PER-EXAMPLE U-MIAs, where a dedicated attacker is instantiated per datapoint. A…
-
Fast Yet Effective Machine Unlearning
A ZERO-GLANCE method: it never looks at the forget samples, only the forget class LABEL plus a small subset of retain data. Learns an error-maximizing noise matrix correlated with the unlearning class label, then does a…
-
Are we making progress in unlearning? Findings from the first NeurIPS unlearning competition
A principled empirical scoring framework for unlearning. Estimates a per-example epsilon from the error rates of the STRONGEST attacker in a swept family of threshold attacks distinguishing unlearned from retrained mode…
-
A Survey of Machine Unlearning
A broad survey of machine unlearning: definitions, taxonomies, methods, evaluation and open problems. Its own framing is that prior work "has not been able to completely solve the problem due to the lack of common frame…
-
When to Forget? Complexity Trade-offs in Machine Unlearning
Not a method paper. Bounds the minimax computation time of unlearning against retraining for strongly convex objectives, via the "unlearning complexity ratio" T^U_e / T^S_e, yielding a phase diagram with three regimes:…
-
Machine Unlearning via Information Theoretic Regularization
Argues the standard anchor-based unlearning principle ("indistinguishable from a model retrained on the retain set") is INAUDITABLE, because it depends on an unknown retrained anchor in parameter space. Replaces it with…
-
Near-Optimal Pure Machine Unlearning for Smooth Strongly Convex Losses
Establishes MATCHING upper and lower bounds on excess population risk for pure (delta = 0) eps-unlearning under smooth strongly convex SCO. The algorithm ("core-swap") interpolates between returning the full-data ERM an…
Nothing matches those filters.
Peripheral
Nothing in the corpus references these. A finding about the corpus, not a defect in it:
izzo2021approximate
Reading state
- partial read 27