Introduction
Federated Learning (FL) is often promoted as a privacy-preserving paradigm because raw data remains on the client’s device. However, this paper introduces Enhanced Source Inference Attacks (ESIAs), demonstrating that a malicious server can actively manipulate the training process to "unmask" which specific client contributed a particular data sample. By combining gradient ascent with data augmentation, the researchers show that an active adversary can identify data origins with significantly higher accuracy than previous passive methods.
The Problem: Identifying Data Sources
In standard FL, clients share model updates (gradients) rather than raw data. While this mitigates direct exposure, it remains vulnerable to Membership Inference Attacks (determining if a record was in the training set). This research addresses a more invasive threat: Source Inference. If a server identifies a sensitive record within the global training set, can it pinpoint the exact hospital or user that provided it?
Existing source inference attacks are primarily "passive," merely observing updates. These often fail because behavioral differences between clients are not sufficiently distinct. This paper explores an "active" threat model where a server intentionally sabotages the global model to force the true data owner to reveal themselves through their local training response.
Methodology: The Trap-and-Trace Strategy
The core of the ESIA approach is a "trap-and-trace" strategy involving two main components:
- Active Gradient Ascent: In each training round, the malicious server takes the target record ($z$) and intentionally performs gradient ascent on the global model. This deliberately increases the model's loss on that specific record, essentially "breaking" the model's ability to recognize that data point.
- Data Augmentation: To ensure robustness, the server targets an "augmented set" of the record (including rotations, flips, and noise) rather than just the original sample.
The Logic: When the "broken" global model is sent to clients, most will continue training on their own unrelated data. However, the source client—who actually owns the target record—will naturally minimize the loss on its local data, thereby "repairing" the damage the server inflicted. By comparing updated models, the server identifies the client showing the lowest loss on the target record and its variants.
Experimental Setup
The authors evaluated ESIAs across diverse benchmarks:
- Image Data: MNIST, FEMNIST, CIFAR-10, and CIFAR-100.
- Tabular/Synthetic Data: A custom Synthetic IID dataset and the Purchase dataset.
- Heterogeneity Control: A Dirichlet distribution (parameterized by $\alpha$) was used to simulate varying levels of Non-IID data, a common characteristic of real-world FL.
Key Results and Insights
ESIAs were tested against five major FL frameworks (FedAvg, FedSGD, FedMD, FedPer, and FedProx) and consistently outperformed existing baselines.
- Superior Accuracy: In a FedAvg scenario with CIFAR-10 ($\alpha=0.1$), ESIA achieved an Attack Success Rate (ASR) of 74.3%, significantly higher than standard passive attacks (62.5%).
- The Impact of Local Training: Increasing the number of local epochs ($E$) improved the ASR. For the Purchase dataset, the ASR rose from 73.2% to 78.3% as local training increased, suggesting that more local optimization makes clients more "visible" to the attacker.
- Non-IID Data as a Risk Factor: While data heterogeneity is a challenge for FL performance, it aids attackers. The more unique a client's data distribution is, the easier it is to pinpoint them as the source of a specific record.
Conclusion
This research serves as a stark reminder that "keeping data on-device" is not a complete privacy solution. Active adversaries can exploit the natural behavior of local optimization to bypass the perceived anonymity of Federated Learning. Future work may involve exploring spatio-temporal features to identify sources in even more complex environments, such as healthcare networks.
Summary updated