More than Enough is Too Much: Adaptive Defenses against Gradient Leakage in Production Federated Learning
Bibliographic record
Abstract
With increasing concerns on privacy leakage from gradients, a variety of attack mechanisms emerged to recover private data from gradients at an honest-but-curious server, which challenged the primary advantage of privacy protection in federated learning. However, we cast doubt upon the real impact of these gradient attacks on production federated learning systems. By taking away several impractical assumptions that the literature has made, we find that gradient attacks pose a limited degree of threat to the privacy of raw data.Through a comprehensive evaluation on existing gradient attacks in a federated learning system with practical assumptions, we have systematically analyzed their effectiveness under a wide range of configurations. We present key priors required to make the attack possible or stronger, such as a narrow distribution of initial model weights, as well as inversion at early stages of training. We then propose a new lightweight defense mechanism that provides sufficient and self-adaptive protection against time-varying levels of the privacy leakage risk throughout the federated learning process. As a variation of gradient perturbation method, our proposed defense, called Outpost, selectively adds Gaussian noise to gradients at each update iteration according to the Fisher information matrix, where the level of noise is determined by the privacy leakage risk quantified by the spread of model weights at each layer. To limit the computation overhead and training performance degradation, Outpost only performs perturbation with iteration-based decay. Our experimental results demonstrate that Outpost can achieve a much better tradeoff than the state-of-the-art with respect to convergence performance, computational overhead, and protection against gradient 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 | 0.008 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.002 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.001 |
| Open science | 0.007 | 0.026 |
| Research integrity | 0.000 | 0.000 |
| 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; both teacher heads agree on what is shown here.
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".