FedUIMF: Unambiguous and Imperceptible Model Fingerprinting for Secure Federated Learning on Consumer Electronic Devices
Bibliographic record
Abstract
The proliferation of consumer electronic devices has generated massive volumes of decentralized data, yet centralized training methods may potentially infringe on privacy. Consequently, federated learning (FL), a novel methodology leveraging distributed collaborative training, has emerged as a viable solution to protect user data privacy. However, the collaborative essence and substantial value of FL models make them vulnerable to theft, endangering the intellectual property (IP) rights of the model owner. While model watermarking is a frequently utilized strategy for protecting FL models, its intrusiveness may compromise the primary task, prompting the exploration of model fingerprinting as an alternative protective measure. Nevertheless, current model fingerprinting methods encounter difficulties in resisting ambiguity attacks, attaining efficiency, and ensuring imperceptibility, rendering them unsuitable for secure FL settings. To address these challenges, this paper proposes FedUIMF, an unambiguous and imperceptible model fingerprinting method for secure federated learning on consumer electronic devices. Initially, clients generate model fingerprints to adapt to the federated learning environment. Leveraging this characteristic, we design a unique identity embedding algorithm that ensures the high uniqueness of the model fingerprint identity through a triple-identity mechanism that integrates information hiding, timestamps, and digital signatures. Additionally, we propose an imperceptible model fingerprinting algorithm that utilizes the Discrete Wavelet Transform (DWT) and the Just Noticeable Distortion (JND) model. By selecting fingerprint samples near the decision boundary, applying DWT to minimize high-frequency disturbances, and guiding perturbations using the JND model, FedUIMF efficiently generates imperceptible model fingerprints. Experimental evaluations conducted on the CIFAR-10, Tiny-ImageNet, and ImageNet-1K datasets revealed that FedUIMF attains exceptional perceptual metric scores of 1.00 (SSIM), 59.29 (PSNR), and 0.0002 (LPIPS), while exhibiting robust model copyright verification with AUC value of 1 in diverse FL settings and under various piracy attacks.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.001 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.004 | 0.000 |
| Research integrity | 0.000 | 0.002 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".