acceptodds
Under review as a conference paper at ICLR 2027

mmSLP: Semantic-Level Privacy Protection for mmWave Human Activity Recognition

Abstract

Millimeter-wave (mmWave) radar enables privacy-friendly human activity recognition (HAR) without visual sensing. However, mmWave-based HAR may still reveal sensitive behaviors, raising privacy concerns for users. Therefore, we aim to reduce the sensitive-activity recognition capability of HAR systems. We observe that HAR systems commonly require user-provided data for fine-tuning, and these data remain under the user’s control before sharing. We therefore apply carefully designed perturbations to these data before fine-tuning to reduce the recognition accuracy of sensitive activities. Based on this idea, we propose mmSLP, a semantic-level privacy protection framework for mmWave HAR. Experiments show that mmSLP reduces sensitive-activity accuracy from 85.41% to 42.12%, while maintaining normal-activity accuracy at 88.09% compared with 84.89% under clean fine-tuning, with low computational overhead.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
mmWave radar, human activity recognition, privacy protection, class-selective protection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_qPCyOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes mmSLP, a semantic-level privacy protection framework for mmWave human activity recognition (HAR) that suppresses sensitive-activity recognition while preserving normal-activity utility. The key idea is to inject class-wise perturbations only into sensitive fine-tuning samples, inducing shortcut learning that degrades recognition of clean sensitive activities. To reduce computational overhead, mmSLP combines error-minimizing and perturbation–target binding objectives with lightweight first-order optimization. Experiments across three datasets and multiple architectures show that mmSLP reduces sensitive-activity accuracy from 85.41% to 42.12%, while maintaining normal-activity with low computational overhead.

Critical strengths
  1. The paper addresses an important privacy concern in mmWave-based HAR by enabling selective protection of user-defined sensitive activities while preserving normal-activity recognition. This fine-grained privacy–utility trade-off is more practical than indiscriminately disabling activity recognition and represents a meaningful research direction.
  2. The proposed method is conceptually straightforward, using class-wise perturbations and a first-order approximation without requiring an additional perturbation generator or higher-order gradient computation. This lightweight design reduces computational and memory overhead while maintaining effective sensitive-activity suppression, making it potentially suitable for resource-constrained user-side deployment.
Critical weaknesses
  1. I notice that the Random Noise baseline uses independently sampled noise for each sensitive sample, which fundamentally differs from the proposed class-wise shared perturbations. The authors should evaluate fixed class-wise random perturbations and a single perturbation shared across all sensitive classes, which may induce shortcut learning without any additional optimization. These comparisons are crucial to establish whether the gains come from the proposed optimization or merely from consistent perturbation patterns.
  2. The perturbation budget η = 0.2 allows an L2 distortion of up to 44.7% of the original signal norm, raising concerns that the observed protection may stem from obscuring the original signals rather than inducing shortcut learning. The authors should evaluate replacing sensitive samples entirely with constant-valued signals as an unconstrained reference to investigate whether the proposed optimization is truly necessary.
  3. The proposed binding objective is closely related to BAIT (ICLR 2026), yet BAIT is only discussed in the related work and not directly evaluated. Recent open-source unlearnable-example methods, such as Versatile Transferable Unlearnable Example Generator (NeurIPS 2025), also provide relevant methodological comparisons. Although these methods were originally developed for different modalities, adapting them to the same radar datasets, sensitive-activity partitions, and downstream fine-tuning protocols would help establish the novelty and effectiveness of mmSLP.
  4. The ablation studies mainly investigate computational efficiency, including generator-based perturbations and first-order versus second-order optimization. However, the paper does not isolate the contributions of the error-minimizing loss, binding loss, and surrogate adaptation.
  5. Appendix Table 6 shows that simply withholding sensitive samples already reduces sensitive-activity accuracy to 52.02%, compared with 42.12% for mmSLP, while achieving slightly better normal-activity accuracy (88.35% vs. 88.09%). The authors should justify whether this additional 9.90 percentage-point reduction warrants the extra computation and disclosure of labeled sensitive data.
Questions for authors
  1. Could the authors evaluate fixed class-wise random perturbations, a single shared perturbation across sensitive classes, and constant-template replacement, to determine whether the proposed optimization is necessary? Additionally, could complete signal replacement be evaluated as an unconstrained reference to distinguish shortcut learning from simple signal destruction?
  2. Could the authors compare mmSLP with adapted implementations of recent closely related methods, particularly BAIT (ICLR 2026) and VTG (NeurIPS 2025), under identical radar datasets, perturbation budgets, and downstream evaluation protocols? If such comparisons are infeasible, could the authors explain the specific technical incompatibilities?
  3. Could the authors provide ablations for (i) error-minimizing loss only, (ii) binding loss only, and (iii) optimization without temporary surrogate adaptation?
  4. The choices of λbindλ_{bind} = 3.5, λemλ_{em} = 0.1, and λnλ_n = 2 lack sufficient justification. Could the authors provide a systematic sensitivity analysis to demonstrate that the results are robust rather than dependent on carefully tuned hyperparameters?
8 comments

We thank the reviewer for recognizing the significance of selective privacy protection and the lightweight design of our method. In response to your suggestions, we have added experiments on shared random perturbations and signal replacement, adapted comparisons with BAIT and VTG, ablation studies of key components, and sensitivity analyses of loss weights. We have also further discussed the additional benefits, computational overhead, and data disclosure trade-offs of mmSLP compared with withholding sensitive data.


[W1 W2 Q1] Response

Thank you for suggesting these controls. We evaluated shared random perturbations and complete signal replacement on XRF55 using ResNet18 with RDRA, RD, and RA inputs. All methods used the same user splits and downstream fine-tuning protocol. We modified only the sensitive samples in the fine-tuning set and evaluated the resulting models on clean test samples.

Shared random perturbations and unconstrained signal replacement on XRF55. Each cell reports normal-activity accuracy (higher is better) / sensitive-activity accuracy (lower is better), both in percent.

MethodRDRARDRA
Independent random perturbation90.38 / 83.7592.13 / 77.0391.56 / 80.31
Fixed class-wise random perturbation90.63 / 84.6992.06 / 75.9491.38 / 76.25
One random perturbation for all sensitive classes91.13 / 85.9491.88 / 76.7291.00 / 79.84
Constant replacement (unconstrained)92.25 / 61.1992.44 / 64.7592.63 / 65.53
Zero replacement (unconstrained)92.31 / 52.0992.75 / 62.0992.50 / 58.34

Shared random perturbations. Sensitive-activity accuracy remained between 75.94%75.94\% and 85.94%85.94\% across the random baselines. Sharing a fixed random perturbation within each sensitive class or across all sensitive classes did not consistently improve protection over independent noise under the same budget cap.

Complete signal replacement. As shown in the table, unconstrained constant and zero replacement yielded mean sensitive-activity recognition accuracies of 63.8%63.8\% and 57.5%57.5\%, respectively, both higher than that of mmSLP. Thus, under the perturbation budget, mmSLP suppressed sensitive-activity recognition more effectively than complete signal replacement.


[W3 Q2] Response

We thank the reviewer for suggesting these comparisons. We adapted both methods to XRF55 by perturbing only sensitive fine-tuning samples and using the same relative perturbation budget cap (η=0.2\eta=0.2) and downstream fine-tuning protocol. For BAIT, we retained its perturbation generator, binding objective, and second-order surrogate optimization. For VTG, we used an adversarial radar gain field for domain augmentation and radar class prototypes in place of CLIP text embeddings for perturbation--label alignment.

Radar adaptations of BAIT and VTG on XRF55. Each cell reports normal / sensitive activity accuracy (%).

MethodRDRARDRA
BAIT91.00 / 64.6992.21 / 58.1392.31 / 61.41
VTG90.17 / 81.8891.25 / 73.9191.56 / 77.19

Across RDRA, RD, and RA inputs, BAIT achieved normal-activity recognition accuracies of 91.00% to 92.31% and sensitive-activity recognition accuracies of 58.13% to 64.69%. For VTG, the corresponding ranges were 90.17% to 91.56% and 73.91% to 81.88%, respectively.

These results suggest that, in the evaluated settings, both adaptations preserved normal-activity recognition accuracy while providing limited suppression of sensitive-activity recognition on clean test samples. In the VTG adaptation, a radar gain field and radar class prototypes replaced the original domain augmentation and label alignment mechanisms, respectively. These changes may affect how well the learned perturbations transfer to downstream models.


[W4 Q3] Response

We thank the reviewer for asking us to isolate the contributions of the two loss terms and temporary surrogate adaptation. In the ResNet18--RDRA setting on XRF55, the full mmSLP configuration in the main table achieved 94.50% normal-activity accuracy and 40.16% sensitive-activity accuracy. EM-only removes the binding loss from the perturbation objective, whereas Binding-only removes the error-minimization loss. The third ablation disables temporary surrogate adaptation while retaining the persistent surrogate updates.

Ablations on XRF55 with ResNet18 and RDRA inputs. Approximate values are provisional. Dashes indicate values pending verification.

SettingNormal accuracy (%) ↑Sensitive accuracy (%) ↓
EM-only93.1954.94
Binding-only92.0648.25
Without temporary surrogate adaptation92.3141.7

The provisional results suggest that the full objective suppresses sensitive-activity recognition more effectively than either loss alone.

[Q4] Response

We thank the reviewer for suggesting a sensitivity analysis. On XRF55 with ResNet-18 and RDRA inputs, we halved and doubled the weights of the binding and error-minimization losses around their default values. Sensitivity to the binding and error-minimization loss weights in the earlier implementation.

Settingλbind\lambda_{\rm bind}λem\lambda_{\rm em}Normal acc. (%) ↑Sensitive acc. (%) ↓
Sweep default3.50.1092.1340.16
Binding weight halved1.750.1092.5040.78
Binding weight doubled7.00.1092.4441.72
Error-minimization weight halved3.50.0592.2541.41
Error-minimization weight doubled3.50.2092.7541.25

Across the tested settings, normal-activity accuracy varied by at most 0.620.62 percentage points and sensitive-activity accuracy by at most 1.561.56 percentage points. The default pair yielded the lowest sensitive-activity accuracy in this sweep, with similar performance at the neighboring settings. For λn\lambda_n, Table~7 in the appendix section \emph{Effect of Normal-Activity Weighting} compares equal weighting (λn=1\lambda_n=1) with the main-table setting (λn=2\lambda_n=2). For ResNet-18, increasing λn\lambda_n from 11 to 22 raised normal-activity accuracy from 93.17%93.17\% to 94.50%94.50\% and lowered sensitive-activity accuracy from 41.56%41.56\% to 40.16%40.16\%. ConvLSTM and ViT showed the same direction of change. Performance was relatively stable when varying λbind\lambda_{\rm bind} or λem\lambda_{\rm em}. The comparison of λn=1\lambda_n=1 and λn=2\lambda_n=2 likewise preserved the overall utility--privacy trade-off, suggesting that the reported results do not depend on a narrowly tuned choice of these weights.


[W5] Response

We thank the reviewer for highlighting this trade-off. We agree that withholding sensitive samples is an effective and low-cost baseline to some extent. It achieves slightly higher average non-sensitive-class accuracy than mmSLP (88.35% vs. 88.09%) and does not require perturbation generation or the sharing of sensitive samples. However, the value of mmSLP lies in providing additional protection when withholding sensitive samples alone does not sufficiently meet users' privacy requirements, by further reducing the recognition accuracy of sensitive activities.

The average reduction of 9.90 percentage points also masks substantially larger improvements in certain settings. For example, on Deep-Soli, mmSLP reduces sensitive-class accuracy from 56.54% to 36.94%, while decreasing non-sensitive-class accuracy by only 1.29 percentage points. With ResNet-18, mmSLP achieves an additional reduction of 29.32 percentage points in sensitive-class accuracy, with only a 0.63-percentage-point decrease in non-sensitive-class accuracy. Furthermore, across all 15 experimental configurations, mmSLP consistently achieves lower average sensitive-class accuracy than fine-tuning with only non-sensitive data, demonstrating the consistency of its additional privacy protection benefits.

Regarding computational overhead, under the reported XRF55 configuration, perturbation generation takes approximately 1.75 minutes per user on average. This process is performed before sharing the fine-tuning data and does not require perturbations to be repeatedly applied to future measurements.


We look forward to your further feedback. Thank you again for your time and consideration, and we wish you all the best.

I thank the authors for their additional experiments and clarifications. However, I still have the following concerns:

  1. For the fixed class-wise random perturbation baseline, I expect one fixed random perturbation per sensitive class, with an L2 norm of up to 44.7% of the original signal norm (similar to η = 0.2), applied to all samples of that class. Similarly, the globally shared baseline should apply one fixed random perturbation to all sensitive classes. Could the authors provide implementation details and confirm whether the mean and variant of random variable follow these settings?
  2. Theoretically, such fixed class-wise random perturbations should be sufficient to induce shortcut learning, since the model only needs to recognize a distinctive random vector to identify the corresponding activity class. However, the reported results show the opposite. Could the authors provide a detailed explanation and further experimental evidence to support their findings?
  3. The additional comparisons with BAIT and VTG largely address my original concern. However, the discussion lacks a direct analysis of mmSLP relative to these methods, particularly an explanation of mmSLP's advantages and methodological contributions.
Reviewer_rD6XOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes mmSLP, a privacy protection method for mmWave-based human activity recognition. It adds perturbations to sensitive training samples to prevent downstream models from learning these activities, while preserving normal activity recognition. The method also uses a lightweight optimization strategy. Experiments on three datasets show reduced accuracy on sensitive activities with little impact on normal activities.

Critical strengths
  1. The method is simple and computationally efficient, making it potentially suitable for user-side deployment.
  2. The experiments cover three datasets and multiple models, with additional transferability and ablation studies.
Critical weaknesses
  1. Unclear practical motivation. The paper reduces sensitive activity recognition accuracy to protect privacy. However, these activities may simply be misclassified as normal ones, leading to incorrect system behavior. It is unclear how this preserves the utility of HAR in real applications.
  2. Limited advantage over withholding sensitive data. Appendix A.3 shows that simply excluding sensitive activities from fine-tuning already provides considerable privacy protection. Although mmSLP performs better, the paper does not clearly explain why sharing perturbed sensitive data is necessary or useful in practice.
Questions for authors
  1. If sensitive activities are misclassified as normal activities, wouldn't this lead to incorrect behavior recognition and affect the actual utility of HAR systems? Could the authors clarify how the proposed method handles this issue in practice?
1 comment

We thank the reviewer for recognizing the simplicity, computational efficiency, and breadth of evaluation of our method. In response to your concerns, we have added an analysis of sensitive-activity misclassification and a rejection mechanism to clarify their implications for practical system utility. We have also clarified the additional protection offered by mmSLP over withholding sensitive data and the scenarios in which it is useful.

[W1 and Q1] Response

We thank the reviewer for raising this question about practical deployment. A deployed HAR system can reduce false activity records and unintended actions by accepting predictions only when their confidence exceeds a threshold. Predictions at or below the threshold can be marked as unconfirmed and excluded from activity logs and action triggers.

Sensitive-to-regular misclassifications can occur in our closed-set evaluation. This setting is designed to evaluate mmSLP and assigns each input the highest-scoring activity label. When a deployed system applies a confidence threshold before acting on a prediction, it can intercept low-confidence misclassifications before they affect the application.

In our supplementary Deep-Soli experiment with ResNet-18, confidence-based rejection reduced sensitive-to-regular misclassifications by 48.92%, from 139 of 180 samples (77.22%) to 71 of 180 (39.44%), while regular-activity accuracy remained at 91.67%, compared with 94.58% without rejection. We used the maximum softmax probability as the confidence score and returned null when the score was at or below the threshold.

Activity recognition on Deep-Soli with and without null rejection.

MetricWithout nullWith null
Sensitive → regular139/180 (77.22%)71/180 (39.44%)
Sensitive → null082/180 (45.56%)
Sensitive activities correctly recognized41/180 (22.78%)27/180 (15.00%)
Regular activity recognition accuracy454/480 (94.58%)440/480 (91.67%)

[W2] Response

We thank the reviewer for the valuable suggestion. We agree that withholding sensitive activity data can provide a certain level of protection. However, the model may still recognize activities that users wish to conceal by leveraging its pretrained knowledge and features shared between sensitive and non-sensitive classes. For example, in Deep-Soli, both sensitive and non-sensitive activities involve local finger movements, hand displacements, or changes in palm orientation. Therefore, simply excluding sensitive-class samples from fine-tuning does not directly prevent the model from exploiting these existing cues. In contrast, mmSLP introduces optimized perturbations to influence the discriminative cues learned during downstream adaptation, thereby further reducing the recognition accuracy of sensitive activities.

This additional suppression is particularly evident in the ResNet-18 configuration on Deep-Soli, where sensitive-class accuracy decreases from 50.80% with non-sensitive-only fine-tuning to 21.48% with mmSLP, representing an additional reduction of 29.32 percentage points. In other words, the proportion of correctly recognized sensitive activities drops from approximately one-half to approximately one-fifth. One possible explanation is that the convolutional architecture of ResNet-18 tends to extract local patterns. Consequently, repeatedly occurring, specifically optimized perturbations within the same class may be more readily exploited as discriminative cues. When these cues are absent from clean test samples, the recognition of sensitive activities may be more substantially suppressed. This benefit is not limited to a single model. Across the three architectures evaluated on Deep-Soli, the average sensitive-class accuracy decreases from 56.54% to 36.94%, representing an additional reduction of 19.60 percentage points.

The additional computational overhead is also modest. Under the reported XRF55 configuration, perturbation generation takes approximately 1.75 minutes per user on average. This process is performed only before sharing the fine-tuning data and does not need to be repeated for subsequent measurements. Therefore, sharing perturbed sensitive data provides an option for users who seek stronger suppression of sensitive activity recognition than withholding sensitive data alone. Furthermore, across all 15 experimental configurations covering three datasets, different model architectures, and radar representations, mmSLP consistently achieves lower average sensitive-class accuracy than fine-tuning with only non-sensitive data, demonstrating that its additional privacy protection benefits are consistent across diverse configurations.


We look forward to hearing your thoughts and would appreciate any further feedback when convenient. Thank you again for your time, and we wish you all the best.

Comments on the paper

2 comments

We thank the reviewer for recognizing the practical relevance of our research objective, the soundness of our experimental setup, and the evidence of cross-model transfer and privacy–utility trade-offs. In response to your suggestions, we have added confusion and recovery analyses, objective ablations, and comparisons with related methods; expanded the reporting of training procedures and result variability; and highlighted the normal-only fine-tuning baseline in the main results.

[W1 Q1] Response

We thank the reviewer for suggesting that we analyze these misclassification patterns and test whether an alternative classifier can recover sensitive activity recognition. Misclassified sensitive samples are spread across several normal activity classes, and samples from different sensitive classes can receive the same predicted label. This overlap suggests that a fixed remapping based solely on predicted labels may be insufficient to recover sensitive activity recognition.

Figure 4 in the revised manuscript shows the confusion matrix for the original ResNet-18 classification head after fine-tuning on mmSLP-protected data, pooled across four Deep-Soli test users. N1--N8 denote normal activity classes, and S1--S3 denote sensitive activity classes.

Of the 180 sensitive query samples, 119 were misclassified, all as normal activities. S1 was most often confused with N3 and N4. Misclassifications for S2 were spread across several normal classes, with N3 being the most frequent. S3 was most often confused with N1 and N4. N3 received misclassified samples from all three sensitive classes, so an N3 prediction alone does not uniquely identify the true sensitive class. The recalls for S1, S2, and S3 were 53.33%, 25.00%, and 23.33%, respectively.

To test whether the fine-tuned features still support sensitive activity recognition, we froze the feature extractor for each user and trained separate linear recovery heads using 1, 2, or 4 clean, correctly labeled support samples per class. We evaluated the original head and all recovery heads on the same held-out query samples. The table reports mean accuracies across the four test users.

Classifier headShots per classNormal (%)Sensitive (%)Gain (pp)
mmSLP original head094.5822.80—
Linear recovery head192.2973.89+51.09
Linear recovery head293.3379.44+56.64
Linear recovery head492.7176.67+53.87

Even with one clean labeled sample per class, the linear recovery head raises sensitive activity accuracy from 22.80% to 73.89%. This shows that the adapted features retain sensitive activity information that a linear classifier can recover.

Under the evaluated sharing protocol, users upload only perturbed versions of sensitive support samples and keep the clean originals on their devices. The learner therefore cannot obtain clean sensitive samples paired with their correct labels through this workflow. The main results quantify suppression of sensitive activity recognition under the evaluated sharing and fine-tuning protocol. The recovery experiment shows that this suppression can weaken substantially if the learner separately obtains even a small number of clean, correctly labeled sensitive samples.


[W2 Q2] Response

We thank the reviewer for suggesting a matched ablation to quantify the contribution of the binding term. On Deep-Soli, we remove the binding term from the mmSLP objective while retaining the error-minimization term and all other experimental settings. Each cell reports normal-activity accuracy followed by sensitive-activity accuracy, both in percent.

MethodResNet-18ConvLSTMViT
mmSLP without binding97.08 / 32.7891.39 / 45.7090.83 / 54.81
Full mmSLP94.79 / 21.4889.72 / 40.7492.50 / 48.60

Adding the binding term reduces sensitive-activity accuracy by 11.30, 4.96, and 6.21 percentage points on ResNet-18, ConvLSTM, and ViT, respectively. Normal-activity accuracy decreases by 2.29 and 1.67 percentage points on ResNet-18 and ConvLSTM, respectively, and increases by 1.67 percentage points on ViT. Under the evaluated fine-tuning protocol, binding provides additional suppression of sensitive-activity recognition beyond error minimization alone on all three backbones, with a reduction in normal-activity accuracy on two of them.

In the appendix of the revised manuscript, we have added details on the PGD-style baseline, including its differentiable classifier-head adaptation, outer objective, and projected updates. We have also expanded the comparison with BAIT to clarify the differences in optimization objectives, perturbation parameterization, and how gradients are handled through surrogate adaptation.

[W3 Q3] Response

We thank the reviewer for requesting more detail on run-to-run variability and downstream fine-tuning. The main results table in the revised manuscript now reports sample standard deviations across three random seeds for both normal- and sensitive-activity accuracy.

We have added three parts to the appendix. First, we specify the fixed downstream fine-tuning configuration for each dataset, including the number of epochs and the learning-rate schedule. We also explain how we select checkpoints based on support-set training accuracy and evaluate them on the held-out query set. Second, we describe the hyperparameters and surrogate update procedure used to generate mmSLP perturbations. Third, we explain how random seeds affect support-set minibatch formation and the sampling of normal-activity examples. We also describe how we average accuracies across test users within each run and then compute sample standard deviations across the three seed-level averages.


[Q4] Response

We thank the reviewer for this suggestion. We have added the Normal-only Fine-tuning baseline to the main results table in the revised manuscript.


We look forward to your further feedback. Thank you again for your time and consideration, and we wish you all the best.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.