DataCook: Exploring Anti-Adversarial Fingerprints for Healthcare Data Usage Control
Organizations: The Chinese University of Hong Kong Shenzhen, Guangdong, China · Swiss Federal Institute of Technology Lausanne (EPFL), Lausanne, Switzerland
Abstract
Purpose: Healthcare AI increasingly relies on high-quality datasets, yet medical data are scarce, costly to curate, and tightly regulated in terms of data access. More critically, once medical data are leaked or accessed without authorization, data owners have limited means to prevent such data from being used to train and deploy downstream models. To address this gap, we propose DataCook, a deployment-stage usage-control framework based on anti-adversarial fingerprints. Materials and Methods: Specifically, DataCook transforms raw samples through a private cooking procedure that embeds anti-adversarial fingerprints. Models trained on the cooked data retain strong utility when evaluated on similarly cooked inputs, while their performance on raw inputs is reduced, making downstream model utility dependent on access to the private cooking function. We evaluate DataCook on 2D, 3D, and high-resolution medical images, as well as structured healthcare tabular data, across multiple model architectures and adaptive attack settings. Results: Across the evaluated settings, DataCook generally preserves authorized model utility while reducing performance on unauthorized raw inputs. The authorization-dependent behavior is observed across multiple medical imaging settings, model architectures, and adaptive attack conditions, supporting the feasibility of the proposed deployment-stage usage-control mechanism. Conclusion: These results demonstrate the feasibility of authorization-dependent model use at deployment, while the effectiveness of the current framework remains dependent on the specific task and evaluation metric.
Figures & tables
| Method | Label | Type | SSIM | CP | |
| Error-Minimum 18 | W/ | Sample | - | ||
| Adversarial Poison 13 | W/ | Sample | - | ||
| Synthetic Perturbations 55 | W/ | Class | - | ||
| Unlearnable Clusters 59 | W/O | Class | |||
| Error-Minimum (Pseudo) | W/O | Sample | |||
| DataCook (Adv) | W/O | Sample |
| Dataset | Task | AUC- CP | AUC- PP |
| MedMNIST 2D | |||
| BloodMNIST | MultiClass | ||
| BreastMNIST | Binary | ||
| ChestMNIST | MultiLabel | ||
| DermaMNIST | MultiClass | ||
| OrganCMNIST | MultiClass | ||
| Dataset | ResNet18 | ResNet50 | VGG16 | ConvNeXt-Tiny | ||||
| CP | CP | CP | CP | |||||
| Path | -71.36 | 0.00 | -47.19 | 2.51 | -65.23 | 1.03 | -76.92 | -0.65 |
| Derma | -25.44 | 0.00 | -10.12 | 1.59 | -7.83 | 2.90 | -11.68 | 3.23 |
| OCT | -43.40 | 0.01 | -49.60 | -3.99 | -30.40 | -2.79 | -10.30 | -3.29 |
| Pneumonia | -21.15 | 0.03 | -26.44 | 0.35 | -16.18 | 0.67 | -8.33 | 1.32 |
| Retina | -6.00 | -1.57 | -3.75 | -0.10 | -45.51 | 3.18 | -2.75 | -4.58 |
| Method | 3D | High Resolution 2D | ||||||||||||
| Fracture | Nodule | Organ | Synapse | Breast | Pneumonia | Retina | ||||||||
| CP | PP | CP | PP | CP | PP | CP | PP | CP | PP | CP | PP | CP | PP | |
| Error-Minimum 18 | -5.83 | – | -6.77 | – | -3.28 | – | -4.83 | – | -13.46 | – | -29.81 | – | -43.75 | – |
| Adversarial Poison 13 | -10.42 | – | -38.39 | – | -5.74 | – | 1.98 | – | -13.46 | – | -29.81 | – | -16.50 | – |
| Synthetic Perturbations 55 | -18.75 | – | -14.84 | – | -14.59 | – | -31.53 | – | -3.86 | – | -2.25 | – | -14.75 | – |
| Unlearnable Clusters 59 | -1.25 | -1.67 | -6.77 | -1.61 | -6.89 | 3.44 | 2.83 | 3.13 | -13.46 | 0 | -29.81 | -2.25 | -23.00 | 0.25 |
| Dataset | Method | Label | CP | PP |
| Pima | Adversarial | W/ | -31.82 | 22.08 |
| Anti-Adversarial | W/ | -21.43 | 22.08 | |
| DataCook (Adv) | W/O | -9.74 | 0 | |
| DataCook | W/O | -14.37 | -0.08 | |
| Hepatitis C | Adversarial | W/ | -72.36 | -1.63 |
| Anti-Adversarial | W/ | -30.89 | 8.13 |
| Dataset | Input | Strongest attack | Before | After | Gain | CP after | Recovery (%) | ||
| BloodMNIST | 28 28 | 8 | 8 | FT-10 | |||||
| BreastMNIST | 28 28 | 8 | 8 | FT-10 | |||||
| ChestMNIST | 28 28 | 8 | 3 | RFT-10 | |||||
| DermaMNIST | 28 28 | 8 | 8 | FT-10 | |||||
| OrganCMNIST | 28 28 | 8 | 8 | JPEG (Q=95) | |||||
| PneumoniaMNIST | 28 28 | 8 | 8 | FT-10 |
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | #Samples | Task | #Labels | Main Features |
| Pima Diabetes | 768 | Binary | 2 | Pregnancies, Glucose, BloodPressure, BMI, Insulin, Age, etc. |
| Hepatitis C | 615 | Multi-Class | 5 | Age, Sex, ALB, ALP, ALT, AST, BIL, CHE, CHOL, CREA, GGT, PROT |
| Systemic Diseases | 61,605 | Multi-Label | 5 | Age, Gender, Ethnic, BMI, Cholesterol, BP, Diabetes, Hypertension, Smoking, Follow-time |
| Parameter | 2D (28 28) | 2D (224 224) | 3D |
| Model Training | |||
| Surrogate training epochs | 100 | 100 | 100 |
| Protected-model training epochs | 100 | 100 | 100 |
| Optimizer | SGD | SGD | SGD |
| Initial learning rate | |||
| Momentum | |||
| Dataset | Task | Task Acc. | Bal. Acc. | AUC | Macro-F1 | Precision | Sensitivity | Specificity | ECE | |||||||||
| MedMNIST 2D | ||||||||||||||||||
| BloodMNIST | MC | 8 | ||||||||||||||||
| BreastMNIST | Bin. | 8 | ||||||||||||||||
| ChestMNIST | ML | 3 | – | – | – | – | ||||||||||||
| DermaMNIST | MC | 8 | ||||||||||||||||
| Metric | Clean | Authorized | Unauthorized | |||
| Subset ACC | ||||||
| Macro-AUC | ||||||
| Micro-AUC | ||||||
| Macro-F1 | ||||||
| Micro-F1 | ||||||
| Macro-Precision |
| Attack | Macro-AUC | Subset ACC | |||||
| After | Gain | CP after | After | Gain | CP after | ||
| Input-transformation attacks | |||||||
| JPEG ( ) | 3 | ||||||
| JPEG ( ) | 3 | ||||||
| JPEG ( ) | 3 | ||||||
| Gaussian ( ) | 3 | ||||||
| Dataset | Input | Task | Pseudo ACC (%) | Confidence (%) | |||
| MedMNIST 2D | |||||||
| BloodMNIST | Multi-class | 8 | |||||
| BreastMNIST | Binary | 8 | |||||
| ChestMNIST | Multi-label | 3 | |||||
| DermaMNIST | Multi-class | 8 | |||||
| OrganCMNIST | Multi-class | 8 | |||||
| Dataset | Input | Pseudo-label variable | with CP-AUC | with PP-AUC | |
| MedMNIST 2D | |||||
| BloodMNIST | Mean confidence | 40 | |||
| BreastMNIST | Pseudo accuracy | 40 | |||
| BreastMNIST | Mean confidence | 40 | |||
| DermaMNIST | Pseudo accuracy | 40 | |||
| DermaMNIST | Mean confidence | 40 | |||
| Dataset | JPEG 50 | JPEG 75 | JPEG 95 | Gauss. 0.5 | Gauss. 1 | Gauss. 2 | Median 3 | Median 5 | Aug. 1 | Aug. 2 | Aug. 3 | ||
| Blood-28 | 8 | 8 | |||||||||||
| Breast-28 | 8 | 8 | |||||||||||
| Chest-28 | 8 | 3 | |||||||||||
| Derma-28 | 8 | 8 | |||||||||||
| OrganC-28 | 8 | 8 | |||||||||||
| Pneumonia-28 | 8 | 8 |
| Dataset | Modality | Primary metric | Clean mean SD | 95% CI | Authorized cooked minus clean | Unauthorized raw minus clean | |||||
| (pp) | Holm | (pp) | Holm | ||||||||
| MedMNIST 2D | |||||||||||
| BloodMNIST | Macro-AUC | 8 | |||||||||
| BreastMNIST | AUC | 8 | |||||||||
| ChestMNIST | Macro-AUC | 3 | |||||||||
| DermaMNIST | Macro-AUC | 8 | |||||||||
| Family or modality | Comparison or scope | Tests | Pairs | Raw | Min. | Min. Holm | Holm |
| Unified task-aware primary-metric tests | |||||||
| 2D | Clean vs. authorized | 31/31 | 3–8 | 30 | 0 | ||
| 2D | Clean vs. unauthorized | 31/31 | 3–8 | 24 | 0 | ||
| 2D | Clean vs. authorized | 3/3 | 3 | 0 | 0 | ||
| 2D | Clean vs. unauthorized | 3/3 | 3 | 0 | 0 | ||
| 3D | Clean vs. authorized | 12/9 | 3 | 0 | 0 | ||
| Dataset | Resolution | SSIM | PSNR (dB) | LPIPS | |||
| MedMNIST 2D | |||||||
| BloodMNIST | 8 | ||||||
| BreastMNIST | 8 | ||||||
| ChestMNIST | 3 | ||||||
| DermaMNIST | 8 | ||||||
| OrganCMNIST | 8 | ||||||
| Dataset | SSIM | PSNR (dB) | LPIPS | ||||
| BloodMNIST | 8 | 2 | |||||
| BloodMNIST | 8 | 4 | |||||
| BloodMNIST | 8 | 8 | |||||
| BloodMNIST | 8 | 12 | |||||
| BloodMNIST | 8 | 16 | |||||
| BreastMNIST | 8 | 2 |
| Dataset | SSIM | PSNR (dB) | ||||
| FractureMNIST3D | 3 | 8 | ||||
| FractureMNIST3D | 3 | 16 | ||||
| FractureMNIST3D | 3 | 32 | ||||
| NoduleMNIST3D | 3 | 8 | ||||
| NoduleMNIST3D | 3 | 16 | ||||
| NoduleMNIST3D | 3 | 32 |
| Dataset | Input | Test | Batch | Cook | Raw infer | Cooked infer | Overhead | GPU (MB) | |
| /sample | |||||||||
| MedMNIST 2D | |||||||||
| BloodMNIST | 3421 | 128 | 1 | ||||||
| BreastMNIST | 156 | 128 | 1 | ||||||
| ChestMNIST | 22433 | 128 | 1 | ||||||
| DermaMNIST | 2005 | 128 | 1 | ||||||
| Dataset | Train | Test | Train cook | Test cook | Test cook/sample | Raw infer | Cooked infer | Protected train | |
| (s) | (s) | (ms/sample) | (ms/sample) | (ms/sample) | (s) | ||||
| Diabetes | 614 | 154 | 0.10 | ||||||
| 0.25 | |||||||||
| 0.50 | |||||||||
| 1.00 | |||||||||
| 2.00 |
Privacy-Preserving ML
Privacy-preserving machine learning develops methods to learn, share, and run models while limiting exposure of sensitive data.
See all 445 papers