Machine Unlearning
Machine Unlearning and Verification Under Black-Box Model Access Constraints
Generated automatically from the limitations stated in 3 papers (ICML, ICLR), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Current machine unlearning algorithms and formal auditing frameworks overwhelmingly require white-box access to model parameters, gradients, or pre- and post-unlearning checkpoint weights. This structurally prevents unlearning operations and compliance verification on closed-source, API-hosted, or third-party proprietary models where internal weights cannot be inspected or directly modified. Without methods and evaluation protocols that operate under black-box or query-only access, unlearning remains unusable for hosted foundation models and unverifiable by external auditors.
Why it matters
Enables external verification of compliance with deletion requests and makes unlearning strategies applicable to proprietary, API-mediated model deployments.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Benchmark black-box unlearning approximations (such as logit suppression, decoding-time intervention, and retrieval-augmented negative prompting) against standard white-box gradient ascent across benchmark forget sets (e.g., ToFU), measuring forget quality and general utility retention.
- 2
Develop and evaluate black-box verification protocols that test whether target concepts have been removed using only input-output interaction and confidence scores, measuring auditing sensitivity without access to model weights or algorithm specifications.
- 3
Quantify the empirical performance gap between white-box parameter edits and query-level black-box interventions across varying model scales and unlearning task granularities (e.g., verbatim sequence removal vs. factual knowledge erasure).
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
Black-box unlearning without parameter updates may reduce entirely to surface-level guardrailing, making it vulnerable to trivial jailbreaks and structurally incapable of matching true weight-space forgetting.
Evidence
Each paper's own statement of the limitation, verbatim.
- Erase or Hide? Suppressing Spurious Unlearning Neurons for Robust UnlearningICLR 2026
Requires white-box access to model gradients and parameters during the unlearning process
- When to Forget? Complexity Trade-offs in Machine UnlearningICML 2025
Assumes the forget set is inaccessible during unlearning; methods that exploit the forget set (and exact/sharding-based unlearning) are excluded from both the framework and the lower bound
- Textual Unlearning Gives a False Sense of UnlearningICML 2025
TULA-DR requires white-box access to weights before and after unlearning and knowledge of the unlearning algorithm, a stronger assumption than the black-box setting
Nearest existing work
- Verification of Machine Unlearning is FragileICML 2024
- Unlearn and Burn: Adversarial Machine Unlearning Requests Destroy Model AccuracyICLR 2025
- Textual Unlearning Gives a False Sense of UnlearningICML 2025
- Probing Hidden Knowledge Holes in Unlearned LLMsNeurIPS 2025
- The Unseen Threat: Residual Knowledge in Machine Unlearning under Perturbed SamplesNeurIPS 2025
- Towards Source-Free Machine UnlearningCVPR 2025
- Unlearning Isn’t Forgetting: Revealing Hidden Leakage in Class Unlearning EvaluationsICML 2026
- Towards Safe Machine Unlearning: a Paradigm that Mitigates Performance DegradationTheWebConference 2025
- Reconstruction Attacks on Machine Unlearning: Simple Models are VulnerableNeurIPS 2024
- Remember What You Want to Forget: Algorithms for Machine UnlearningNeurIPS 2021
- Towards Unbounded Machine UnlearningNeurIPS 2023
- Keeping an Eye on LLM Unlearning: The Hidden Risk and RemedyNeurIPS 2025
- TAPE: Tailored Posterior Difference for Auditing of Machine UnlearningTheWebConference 2025
- Rethinking Adversarial Robustness in the Context of the Right to be ForgottenICML 2024
- What makes unlearning hard and what to do about itNeurIPS 2024
Related open problems
Machine Unlearning
A Standardized Multi-Architecture Testbed of Memorized Concepts, Generation Fingerprints, and Editable Structures in Text-to-Image Models Beyond Stable Diffusion v1.x
Nearly every quantitative claim about text-to-image models — memorization, unlearning success, watermark robustness, deepfake provenance, concept editability — is anchored to a handful of frozen Stable Diffusion v1.x/v2 checkpoints, because those are the only models with established benchmarks, known memorized prompts, and validated concept labels. Newer architectures (DiT-based diffusion, DeepFloyd, autoregressive and GAN generators, video models) are largely untested, so results that appear robust may be artifacts of one 2022 model family. The field cannot tell whether its findings are properties of generative modeling or properties of Stable Diffusion 1.4.
Machine Unlearning
Benchmarking the Adversarial Robustness and Reversibility of LLM Unlearning
Existing LLM unlearning methods are predominantly evaluated on standard, benign queries, giving a false sense of compliance with privacy and copyright demands. Empirical evidence shows that "unlearned" knowledge remains extractable via adversarial jailbreaks, latent-space elicitation, and few-shot relearning. Without systematic evaluation across these extraction vectors, practitioners have no way to verify whether a model has actually eliminated sensitive data or merely applied a superficial suppression mask.
Machine Unlearning
Empirical Robustness and Sensitivity of Machine Unlearning Under Realistic Data and Checkpoint Constraints
Existing machine unlearning algorithms for language models are predominantly evaluated under idealized conditions: complete access to forget sets, clean entity anchors, standard prose, and checkpoints saved immediately before target data exposure. In practice, unlearning requests frequently present partial forget data, domain variations like code, noisy or alias-heavy entity mentions, and checkpoints separated from target exposure by billions or trillions of tokens. Practitioners currently cannot predict whether an unlearning method that succeeds on curated benchmarks like TOFU will retain any efficacy when deployed under these real-world data and provenance constraints.
Machine Unlearning
Machine Unlearning Under Degraded Operational Preconditions
Existing machine unlearning methods rely on strict operational preconditions: access to clean retain datasets ($D_r$), full white-box parameter access, historical pre-training checkpoints, or original pre-unlearned calibration weights. In production environments, intermediate training checkpoints are routinely deleted to save storage, retain data is often inaccessible due to data governance policies, and deployment interfaces may restrict full weight access. Without methods that function in the absence of these preconditions, deployed models cannot legally or practically comply with data deletion requests.