machine unlearning

Existing debiasing methods lack an effective evaluation of whether models can still functionally exploit removed label-attribute shortcuts. This work proposes Association Recovery Testing (ART), which assesses debiasing efficacy from a functional perspective by estimating the direction of attribute associations conditioned on class labels, amplifying residual feature signals, and evaluating the modified features using the original classification head to detect recoverable shortcuts. ART reveals residual shortcuts that conventional output-based metrics and representation probes fail to capture, thereby shifting the focus of debiasing approaches from individual classes or isolated concepts toward relational associations. Experiments on Waterbirds, CelebA, SpuCoDogs, and an extended ISIC dataset with timestamp artifacts demonstrate that ART effectively uncovers functional shortcuts persisting after state-of-the-art debiasing interventions.

Conventional machine unlearning methods for continual learning (CL) rely on access to historical training data, leading to catastrophic forgetting of previously learned tasks and task recurrence. Method: This paper proposes UnCLe, a data-free unlearning framework that integrates unlearning directly into the CL paradigm. UnCLe employs a hypernetwork conditioned on task embeddings to generate task-specific parameters and introduces a noise-alignment strategy to enable selective, data-agnostic unlearning—without requiring original training data. Contribution/Results: By decoupling unlearning from data dependency, UnCLe preserves forward transfer capability while effectively mitigating knowledge interference and backward task reactivation. Extensive evaluations on multiple visual benchmarks demonstrate that UnCLe supports joint optimization of incremental learning and unlearning across multiple rounds, significantly improving model stability and post-unlearning accuracy retention compared to state-of-the-art baselines.

Diffusion models often inadvertently erase benign concepts that co-occur with harmful content during unlearning—for example, suppressing the concept of “people” when removing nude images—thereby degrading generative capabilities. This work proposes ReCARE, a framework that formally defines and quantifies such co-occurring and retainable entities (CARE). ReCARE automatically constructs a CARE lexicon and introduces a constrained fine-tuning strategy during targeted unlearning to explicitly preserve these essential concepts. We also introduce the CARE score as a dedicated evaluation metric. Experiments across diverse targets—including nudity, Van Gogh style, and tilapia—demonstrate that ReCARE significantly outperforms existing methods in terms of harmful concept removal efficacy, model generalization, and preservation of CARE concepts.

To address the high retraining cost and verification difficulty in machine unlearning, this paper proposes a lightweight, verifiable neural network patching framework for forgetting. Methodologically, it pioneers a provably correct forgetting theory grounded in the “repair” paradigm; introduces a minimal-patch strategy coupled with a gradient-guided iterative selection of representative samples to ensure strict forgetting guarantees while enhancing scalability; and integrates neural network repair, SMT-based satisfiability solving, and efficient patch optimization—enabling both single-point and batch data deletion without retraining. Experimentally, on multi-class benchmarks, the approach achieves high-fidelity forgetting (accuracy drop <1.2%), reduces verification overhead by 3–5×, and significantly lowers memory consumption compared to retraining and fine-tuning baselines.

This paper addresses the challenges of “exact removal of specific training samples” in machine unlearning—namely, high computational overhead from retraining, significant latency, and degradation in model performance. To this end, we propose the Sequence-Aware Sharded and Stratified Training (S3T) framework. S3T employs hierarchical sequential training, disjoint partitioning of data subsets, and layer-wise parameter isolation, enabling theoretically rigorous, zero-loss exact unlearning via deactivation of only affected layers. It is the first method to support high-concurrency deletion requests while guaranteeing zero service interruption. Integrated with parameter-efficient fine-tuning (PEFT) and multi-sequence joint optimization, S3T achieves substantial improvements across multiple benchmarks: 92% reduction in deletion latency, <0.3% accuracy loss, 100% service availability, and formal theoretical guarantees on deletion equivalence and performance consistency.

This work addresses two critical challenges in sequential machine unlearning: the progressive degradation of accuracy on retained data—termed knowledge erosion—and the unintended recovery of previously forgotten samples, known as unlearning reversal. To tackle these issues systematically, the authors propose SAFER, a novel framework that enforces representational stability for retained data while introducing negative logit margin regularization for forgotten data. This dual mechanism establishes a new paradigm that simultaneously ensures effective unlearning and model stability. Experimental results demonstrate that SAFER significantly mitigates both knowledge erosion and unlearning reversal across multiple unlearning rounds, maintaining high accuracy on retained data while effectively erasing information from forgotten samples.

This work addresses a critical gap in current machine unlearning methods, which predominantly rely on output-layer metrics and fail to discern whether sensitive information is genuinely erased or merely suppressed at the representation level. To resolve this ambiguity, the authors propose a recovery-based analytical framework that, for the first time, distinguishes between “suppression” and “deletion” of information in intermediate representations. By leveraging sparse autoencoders to identify class-specific features in hidden layers and applying feature-level interventions during inference, the framework quantitatively measures residual information. Evaluations across 12 state-of-the-art unlearning methods on image classification tasks reveal that most—including fine-tuning of pretrained models—only suppress rather than delete semantic features, as their representations remain highly recoverable. This finding exposes a significant blind spot in existing unlearning evaluation protocols and establishes a new criterion for assessing privacy guarantees.

This work addresses the longstanding conflation in machine unlearning research between “untraining” and “unlearning,” which has led to ambiguous problem formulations and inadequate evaluation criteria. We formally distinguish these concepts for the first time: untraining aims to remove the influence of specific training samples, whereas true unlearning requires erasing the model’s knowledge of the entire underlying data distribution or concept those samples represent. Through theoretical formalization and a systematic review of existing literature, we establish a clear conceptual framework, reclassify current methods accordingly, and uncover critical challenges that have been overlooked. By clarifying foundational definitions, this study lays the groundwork for rigorous algorithmic evaluation, promotes standardization in the field, and delineates promising directions for future research.

This work addresses a critical limitation in current machine unlearning methods, which often suppress rather than truly erase target knowledge, leaving it potentially recoverable. Due to the opacity of large models, verifying the effectiveness of such methods remains challenging. To overcome this, the study implements mainstream unlearning approaches within a fully interpretable two-layer neural network framework, enabling direct analysis of concept representations encoded in model weights. The authors systematically evaluate whether these methods genuinely remove targeted knowledge and assess its recoverability after fine-tuning. Their findings reveal, for the first time in an interpretable setting, that most unlearning techniques merely obscure knowledge rather than eliminate it—“forgotten” information can be efficiently restored through simple fine-tuning, thereby exposing the fragility and fundamental limitations of existing unlearning mechanisms.

This work addresses the challenge of targeted machine unlearning in large language models (LLMs). Unlike conventional fine-tuning or data-deletion approaches, we propose an efficient unlearning method grounded in the model editing paradigm. We are the first to systematically evaluate and adapt causal mediation–based editing algorithms—including ROME, IKE, and WISE—for machine unlearning tasks. Crucially, we reformulate the editing objective to emphasize precise knowledge localization and controllable, localized parameter modification—driven by gradients or activations—enabling accurate removal of targeted information. On multiple standard unlearning benchmarks, our method substantially reduces residual memory rates while limiting downstream task performance degradation to under 3%; in certain settings, it outperforms state-of-the-art unlearning baselines. Our core contributions are: (i) establishing model editing as a novel, high-fidelity paradigm for targeted unlearning; and (ii) introducing principled design criteria and technical pathways for unlearning-aware editing objectives.