acceptodds
Under review as a conference paper at ICLR 2027

Forgotten or Just Unseen? Semantic Erasure and Auditing for Face Unlearning.

Abstract

In open-set face recognition, forgetting an identity and never having seen it in the first place can exhibit indistinguishable behavior, raising a fundamental question of what it truly means to “unlearn” an identity. We term this the Forgetting-Or-Generalization (FOG) ambiguity. Existing unlearning measures rely on accuracy-based metrics, whereas face recognition operates in an open-set, verification-based setting where unseen identities are expected to be recognized correctly. This mismatch makes facial erasure difficult to evaluate and renders verification-accuracy-based unlearning metrics misleading. We formalize this facial unlearning problem and introduce a generative auditing framework for semantic visualization of identity erasure, revealing that existing unlearning methods retain identity information despite appearing successful under conventional metrics. To address this, we propose a feature-space unlearning algorithm that removes structural identity groupings by erasing identity regions using region-aware surrogate mechanisms while preserving recognition of non-target identities. Our framework shifts face unlearning evaluation from accuracy-driven to semantics-based assessment, providing stronger empirical privacy validation. Experiments on CASIA-WebFace, WebFace4M, MS1Mv3, and D-LORD show that FUSE achieves the lowest forget-set semantic residual among nine unlearning methods (as low as 0.27), while preserving retain-set verification and open-set accuracy (99.31% on LFW, against 99.53% for the original model).

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Privacy, unlearning, Face recognition

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_EKNyOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies identity unlearning in open-set face recognition. Since face recognition models are expected to generalize to unseen identities, forgotten identities may remain verifiable. Verification accuracy therefore cannot reliably distinguish successful unlearning from residual training information. The authors call this the Forgetting-or-Generalization (FOG) ambiguity. The authors introduce Semantic Residual (SR), which uses the noise prediction error of a conditional diffusion model to assess identity information that remains usable in embeddings after unlearning. They also propose FUSE, which combines surrogate representations derived from facial regions, identity cone modeling, and geometric constraints to modify target identity representations while preserving the structure of retained identities. Experiments on four face datasets compare FUSE with several existing methods and suggest a favorable balance between unlearning and retaining model utility.

Critical strengths
  1. The paper identifies a meaningful evaluation problem in open-set identity unlearning. Face recognition models need to generalize to unseen identities, so continued verification of a forgotten identity does not necessarily indicate failed unlearning. Likewise, a decrease in verification accuracy does not establish that its training influence has been removed. The discussion of FOG clearly explains this limitation and motivates SR as a complementary evaluation measure.
  2. FUSE shows competitive empirical performance. The reported results indicate that FUSE can reduce semantic residuals for target identities while retaining reasonably strong verification performance on retained and unseen identities.
Critical weaknesses
  1. SR may be sensitive to changes in the feature representation. The primary auditing generator is adapted to the original feature distribution, so a decrease in SR may also reflect a mismatch between the generator and the modified embeddings. The existing experiments do not fully rule out this explanation, and the inversion attack results are also evaluated primarily using SR.
  2. The correlation analysis across generators needs clarification. Table 7 does not specify the correlation measure, unit of analysis, or range of data included. It therefore remains unclear whether these correlations support the claim of consistent method rankings.
Questions for authors
  1. Could you provide a control experiment to distinguish reduced identity information from the generator’s inability to use the modified embeddings? For example, align the new embeddings to the original space using only retained identities, then reevaluate SR and identity recovery.
  2. What SR data, correlation measure, and unit of analysis were used to compute the coefficients in Table 7?
  3. Does “classifier-level changes” in Appendix D mean updating only the classifier head, or updating the encoder as well? If the encoder and the conditioning embeddings used for SR remain unchanged, SR would also be expected to remain unchanged under the same auditing settings.
4 comments

We thank the reviewer for the detailed review and for suggesting a concrete control experiment. We respond to each point below.

W1 and Q1: control for generator mismatch

We agree with the concern. Table 1 already hints at it: Retrain has seen all retained identities, yet its retain residual is 0.650 and not 1. So part of any drop in SR comes from the change in the embedding space. We ran the control you proposed. For each unlearned model, we learn an orthogonal (Procrustes) map from its embedding space to the original space. The map is fitted only on retained identities and never sees forget data. We fit it on 60% of the retained identities and compute the aligned residuals on the remaining 40%. We then recompute SR on the mapped embeddings. We use an orthogonal map because it only rotates the space, so it cannot add or remove identity information.

For identity recovery, we report Rank-1 identification of the aligned forget embeddings against the original prototypes of all 10,000 training identities. This check does not use any generator. As in Table 1, a single identity is unlearned at a time. Rank-1 is averaged over 10 such identities.

MethodRaw GrG_rRaw GfG_fAligned GrG_rAligned GfG_fAligned Gr−GfG_r-G_fRank-1 (%)
OG-----89.5
Retrain0.6500.3350.690.410.2832.7
FT0.4200.3850.570.450.1238.9
GA0.4880.6430.590.62-0.0369.4
L1-sparse0.5400.6480.560.66-0.1058.8
RL0.5590.5390.620.580.0461.6
BS0.5320.5720.570.60-0.0357.8
BE0.5610.5370.620.64-0.0270.4
SalUn0.4980.4870.630.590.0466.5
SG0.5250.4610.630.610.0268.5
FUSE (ours)0.6320.3910.680.420.2637.6
CASIA-WebFace, ResNet-50, ArcFace. Raw values are from Table 1. OG is the original model. Retrain is the exact-unlearning reference.
  • Alignment increases the retain residual for every method. This confirms that raw SR is affected by the change in embedding space, as you expected.
  • FUSE is affected little. Its forget residual moves from 0.391 to 0.42, almost the same as Retrain (0.41).
  • After alignment, FUSE and Retrain are the only methods with a clear gap between retain and forget residuals (0.26 and 0.28). For most baselines the gap is close to zero.
  • Some baselines change more. SG moves from 0.461 to 0.61, BE from 0.537 to 0.64, and SalUn from 0.487 to 0.59. Their raw SR should be read with care.
  • Rank-1 identification agrees. FUSE (37.6%) is close to Retrain (32.7%), while most baselines are between 58% and 70%. FT is also low (38.9%), but its gap is small (0.12).
  • Retrain itself is at 32.7%, far above chance (0.01%). This is the FOG ambiguity, and it is why we compare against Retrain and not against zero.

Revision. This experiment is added to Section 4 and Appendix E.7. The paper now reads SR relative to the retain residual and to Retrain, and lists the dependence on the generator, and the fact that the control is so far on CASIA-WebFace, as limitations in Appendix F.2.

W1 (continued): inversion attack evaluated without SR

We ran a new inversion attack and evaluated it without SR.

Attack. We use an unconditional diffusion model trained on WebFace4M. At each denoising step, we guide it to match the mean embedding of the forgotten identity under the unlearned model. This attack is different from the optimization-based attack in Appendix E.6.1, which we keep.

Evaluation. We embed the generated images with an independent ResNet-101 trained on WebFace4M. This model is not used in training, unlearning, or SR computation. We use 50 forget identities and generate 10 images per identity (500 per method). The gallery has 5,284 CASIA-WebFace identities, each represented by the mean embedding of its real images. A generated image counts as recovered if its nearest gallery identity is the forgotten one and their cosine similarity is at least 0.4. Chance is about 0.02%. The 50 identities are unlearned together in one model per method. This differs from Table 1 and Part 1/4, where a single identity is unlearned.

We also verified that the forgotten identities are not part of WebFace4M. We searched WebFace4M with ArcFace similarity for each of them and manually checked the top-10 matches.

MethodIdentity recovery (%)
Real forget images (upper bound)93.2
Unguided samples (lower bound)0.8
OG85.4
Retrain32.8
FT47.4
GA55.0
L1-sparse48.6
RL52.8
BS49.8
BE57.2
SalUn60.0
SG46.2
FUSE (ours)40.6
  • The attack works. On the original model it recovers the identity almost as often as real images do (85.4% vs. 93.2%). Unguided samples stay near zero (0.8%), so the recovery comes from the attacked model and not from the diffusion prior.
  • FUSE has the lowest identity recovery among the approximate methods (40.6%). The baselines are between 46.2% and 60.0%. The difference to SG, the closest baseline, is small (5.6 points).
  • Retrain is at 32.8%. FUSE covers 85% of the distance from the original model (85.4%) to Retrain.
  • Retrain itself is far above the lower bound. This is again the FOG ambiguity: an identity that was never used in training can still be partly recovered.
  • The attack agrees with SR on the main result, although the two settings differ (50 identities here, one identity in Table 1): FUSE has the lowest independently judged recovery among the approximate methods here, and the aligned SR closest to Retrain in Part 1/4. The order of the baselines is not identical, so we do not claim a ranking among the baselines.

Revision. This experiment is added to Appendix E.6.2, with a summary in Section 4. We also corrected the text of Appendix E.6.1 to match its table.

W2 and Q2: the correlation analysis in Table 7

Thank you for pointing this out. The full specification is below. We first clarify the setting. Table 7 was not computed from the values in Table 3. Table 3 uses a single forget identity, as in Table 1. This reflects a typical deletion request. Table 7 comes from a separate D-LORD experiment in which five identities are unlearned together. The revised captions now state this.

  • Unit of analysis: the unlearning method (9 methods).
  • Data: one GrG_r and one GfG_f value per method and generator, listed in the first table below.
  • Measure: Pearson correlation between two generators, with GrG_r and GfG_f pooled (n = 18).
MethodIDiff-Face GrG_rIDiff-Face GfG_fArc2Face GrG_rArc2Face GfG_fGuided Diff. GrG_rGuided Diff. GfG_f
FT0.4080.4200.4260.3870.3760.433
GA0.4280.3580.4280.3510.3760.410
L1-sparse0.5010.4570.4890.3940.4550.381
RL0.4870.4590.4760.4030.5020.460
BS0.4410.3890.4590.3620.4280.412
BE0.5220.3750.5100.4020.5900.370
SalUn0.5420.4560.5170.4220.5120.448
SG0.4990.3340.5100.3580.4780.384
FUSE (ours)0.5920.3100.6030.3170.5880.322

Pooling GrG_r and GfG_f can inflate the correlation, so we now also report them separately (n = 9 each). Spearman rank correlation is in parentheses.

Generator pairPooled (n = 18)GrG_r onlyGfG_f only
IDiff-Face vs. Arc2Face0.940.96 (0.97)0.85 (0.82)
IDiff-Face vs. Guided Diffusion0.870.90 (0.88)0.74 (0.67)
Arc2Face vs. Guided Diffusion0.840.87 (0.87)0.67 (0.55)

Agreement is high on the retain set and moderate on the forget set. With a single forget identity (Table 3), the order of the baselines on the forget set varies more between generators (Spearman 0.14 to 0.58), while FUSE remains the lowest under all three.

Revision. In the revised PDF, Table 3 is now Table 4 and Table 7 is now Table 8. Their captions state the setting and the measure, Appendix E.4.4 includes both tables above (Tables 8 and 9), and the claim in Section 4 is reworded.

Q3: the meaning of "classifier-level changes"

Thank you for the question. The encoder is updated in all baselines, and we have clarified this in Appendix D. For the classification-based baselines (GA, L1-sparse, RL, BS, BE, SalUn, SG), the loss is computed through the classifier head, as in the original methods. The gradients flow through the whole network, so both the head and the encoder are updated. FT also fine-tunes the whole network. For evaluation we drop the head and use only the encoder's embeddings. These are also the embeddings used for SR. You are right that if only the head were updated, SR would stay at 1. This is not what we see. Baseline SR values in Table 1 are between 0.34 and 0.65, and forget-set verification also changes. For example, GA reduces Vf from 0.836 to 0.518 on CASIA-WebFace.

Revision. Appendix D now reads: "For classification-based baselines (GA, ℓ1\ell_1-sparse, RL, BS, BE, SalUn, SG), the unlearning objective is computed through the classifier head, while all network parameters, including the encoder, are updated; FT also fine-tunes the full network. Evaluation uses only the updated encoder's embeddings, which are also the embeddings used for SR."

We hope these responses address the reviewer's concerns. We are happy to clarify any remaining points.

Reviewer_8vEzOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

summary:

This paper studies identity unlearning in open-set face recognition, where forgotten identities may still be recognized due to the model's generalization ability. The paper introduces the Forgetting-Or-Generalization (FOG) problem and proposes Semantic Residual (SR) to evaluate remaining identity information. It further presents FUSE, a geometry-aware unlearning method that removes forgotten identities from their original feature regions while preserving the recognition ability of other identities. Experiments across four datasets show promising results compared with existing unlearning methods.

Critical strengths

strength:

The paper highlights an important challenge in face unlearning, where verification accuracy alone cannot reliably indicate whether an identity has been forgotten. The FOG formulation provides a useful perspective on this problem and motivates the proposed semantic evaluation.

FUSE provides a reasonable approach to balancing identity forgetting and recognition performance. The experiments cover multiple datasets and model architectures, with comparisons against exact retraining and existing unlearning methods. The additional analyses of SR and identity recovery provide further support for the evaluation.

Critical weaknesses

weakness:

The diffusion model is trained using embeddings from the original recognition model, while the unlearned and retrained models may produce different embedding distributions. This raises the question of whether a lower SR always reflects reduced identity information or partly results from the generator's difficulty in interpreting these embeddings.

The paper argues that forgotten identities should behave like identities never seen during training. However, the evaluation does not consider whether unlearning introduces additional false matches between forgotten and retained identities. Since FUSE moves forgotten embeddings away from their original regions, they may become closer to other identities. This could undermine the claimed unseen-like behavior after unlearning.

There are several unresolved references, such as "Section ??" and "Figure ??". The spherical interpolation formula in Eq. (2) also appears inconsistent with the standard formulation. These issues make some technical details difficult to follow.

After unlearning, is the false match rate between forgotten and retained identities comparable to that between unseen and retained identities?

Questions for authors

see weaknesses

3 comments

We thank the reviewer for the careful review and the helpful questions. We respond to each point below.

W1: does a lower SR mean removed identity information, or only a changed embedding space?

We agree with the concern. Table 1 already hints at it: Retrain has seen all retained identities, yet its retain residual is 0.650 and not 1. So part of any drop in SR comes from the change in the embedding space. To separate the two effects, we ran a control. For each unlearned model, we learn an orthogonal (Procrustes) map from its embedding space to the original space. The map is fitted only on retained identities and never sees forget data. We fit it on 60% of the retained identities and compute the aligned residuals on the remaining 40%. We then recompute SR on the mapped embeddings. An orthogonal map only rotates the space, so it cannot add or remove identity information. We also report a check that needs no generator: Rank-1 identification of the aligned forget embeddings against the original prototypes of all 10,000 training identities. As in Table 1, a single identity is unlearned at a time. Rank-1 is averaged over 10 such identities.

MethodAligned GrG_rAligned GfG_fGr−GfG_r-G_fRank-1 (%)
OG---89.5
Retrain0.690.410.2832.7
FT0.570.450.1238.9
GA0.590.62-0.0369.4
L1-sparse0.560.66-0.1058.8
RL0.620.580.0461.6
BS0.570.60-0.0357.8
BE0.620.64-0.0270.4
SalUn0.630.590.0466.5
SG0.630.610.0268.5
FUSE (ours)0.680.420.2637.6
CASIA-WebFace, ResNet-50, ArcFace. OG is the original model. Retrain is the exact-unlearning reference.
  • After alignment, FUSE has a forget residual of 0.42, almost the same as Retrain (0.41). Before alignment it was 0.391 (Table 1), so the change is small.
  • FUSE and Retrain are the only methods with a clear gap between retain and forget residuals (0.26 and 0.28). For most baselines the gap is close to zero.
  • Some baselines change more under alignment. SG moves from 0.461 to 0.61 and SalUn from 0.487 to 0.59, so their raw SR should be read with care.
  • Rank-1 identification gives the same picture. FUSE (37.6%) is close to Retrain (32.7%), while most baselines are between 58% and 70%. FT is also low (38.9%), but its gap between retain and forget residuals is small (0.12).

Revision. This experiment is added to Section 4 and Appendix E.7. The paper now reads SR relative to the retain residual and to Retrain, and not as an absolute number.

W2 and the question on false matches with retained identities

Thank you for raising this. We have added this check to the paper. In short, the answer is largely yes for FUSE. Averaged over 10 forget identities, each unlearned singly, forgotten probes are falsely matched to retained identities slightly more often than unseen probes (ratio 1.40). This is close to Retrain (1.17) and well below most baselines.

Setup. We compare forget probes against the prototypes of retained identities and compute the false match rate (FMR). We do the same for probes of unseen identities. If forgotten identities behave like unseen ones, the two rates should be similar, so their ratio should be close to 1. All embeddings come from the model under evaluation. The threshold τ\tau is set per model at FMR = 10−310^{-3} on retain impostor pairs. We also report the FMR against the 10 nearest retained classes of each forget identity, where false matches are most likely. As in Table 1, a single identity is unlearned at a time, and the rates are averaged over 10 such runs.

MethodForget →\rightarrow retainUnseen →\rightarrow retainRatioForget →\rightarrow 10 nearest retain
OG1.11.30.852.1
Retrain1.41.21.173.7
FT2.31.61.447.7
GA6.42.13.0510.2
L1-sparse5.82.22.648.2
RL7.31.54.879.5
BS4.82.91.667.1
BE5.11.73.008.9
SalUn6.01.34.629.1
SG5.81.93.058.5
FUSE (ours)2.11.51.404.2
CASIA-WebFace, ResNet-50, ArcFace. All FMR values are ×10−3\times 10^{-3}. Ratio = FMR(forget →\rightarrow retain) / FMR(unseen →\rightarrow retain).
  • FUSE has a ratio of 1.40, slightly higher than Retrain (1.17). Several baselines are much higher: GA, BE, SG, SalUn and RL are between 3.0 and 4.9. These methods do push forgotten embeddings toward retained identities.
  • On the 10 nearest retained classes, FUSE (4.2) is again close to Retrain (3.7). All baselines are between 7.1 and 10.2.
  • FT has a ratio similar to FUSE (1.44), but its FMR on the nearest classes is almost twice as high (7.7). This is consistent with the neighborhood analysis in Appendix E.5, where FUSE preserves 70.1% of the retained k-NN neighborhoods against 37.8% for SalUn.

Revision. This analysis is added to Section 4, with the protocol in Appendix E.5. Section 2 now also states that a forgotten identity should match retained identities about as often as an unseen identity does.

W3: unresolved references and Eq. (2)

Thank you for the careful reading. These are fixed in the revised PDF.

  • "Section ??" in Section 4 and in Appendix E.1 points to Section 2. "Figure ??" in the Class Alignment paragraph points to Figure 6.
  • Eq. (2) had a typo in the first coefficient. The correct form is zicombo=sin⁡((1−β)θ)sin⁡θz1,ireg+sin⁡(βθ)sin⁡θz2,ireg,θ=arccos⁡(z1,ireg⊤z2,ireg).z_i^{\text{combo}} = \frac{\sin\big((1-\beta)\theta\big)}{\sin\theta} z_{1,i}^{\text{reg}} + \frac{\sin(\beta\theta)}{\sin\theta} z_{2,i}^{\text{reg}}, \qquad \theta = \arccos\big(z_{1,i}^{\text{reg}\top} z_{2,i}^{\text{reg}}\big).

Our code uses this standard form, so the results are not affected. In our experiments, β\beta is picked at random between 0.2 and 0.8. While fixing these, we also removed a misplaced sentence after Eq. (5), used the name "Class Alignment" consistently in the text, and reduced Appendix B to a pointer to Appendix E.4.3, which covers the same variant.

We hope these responses address the reviewer's concerns. We are happy to clarify any remaining point.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary rebuttal

We thank both reviewers for their careful and constructive reviews. We respond to each reviewer separately below. This comment summarizes the changes.

The main common concern was whether a lower Semantic Residual (SR) means that identity information was removed, or only that the generator cannot read the changed embeddings. We agree that this needed a control. We ran the alignment experiment suggested by Reviewer EKNy. The conclusion holds: after alignment, FUSE remains close to Retrain on both retain and forget residuals, and no baseline does. Changes in the revised PDF (marked in red):

  • Section 4 has a new paragraph on SR after alignment, with results in the new Table 2. Details are in the new Appendix E.7.
  • Section 4 also reports false match rates between forgotten and retained identities, in the same table. The protocol is in Appendix E.5.
  • Section 4 and Appendix E.6.2 add a new inversion attack that is evaluated with an independent face recognition model.
  • The correlation analysis of the original Table 7 is now fully specified in the revised Table 8, with its input values in Table 9. The related claim in Section 4 is now stated more precisely.
  • Appendix D states which parameters each baseline updates, and the forget-set size of each experiment.
  • We corrected Eq. (2), the unresolved references, and some inconsistent naming. The abstract now reports the gap between retain-set and forget-set SR.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.