Abstract
Federated Learning (FL) is designed to preserve privacy by keeping training data on local devices. However, recent research has exposed a critical vulnerability: malicious participants can exploit shared model updates to reconstruct private training data using Generative Adversarial Networks (GANs). This paper introduces InferGuard, a novel Byzantine-robust aggregation rule designed to mitigate these client-side inference attacks. By utilizing a coordinate-wise median and a distance-based filtering threshold, InferGuard identifies and excludes privacy-leaking updates without requiring a server-side root dataset or compromising model utility.
Introduction
In standard Federated Learning, clients send local gradients to a central server, which aggregates them to update a global model. While this prevents direct data sharing, the GAN-based inference attack allows an adversary to craft local updates that force the global model to reveal information about other clients' data distributions. Essentially, the attacker transforms the global model into a generator that recreates the victim's private images. Existing defenses, such as Differential Privacy or root-dataset-dependent verification, often suffer from significant accuracy degradation or impractical deployment requirements.
Methodology: The InferGuard Defense
InferGuard operates on the observation that malicious updates intended for data reconstruction deviate significantly from the consensus of benign updates. The defense mechanism follows a three-step process in each training round:
- Coordinate-wise Median Calculation: The server calculates the median of all received updates. This serves as a robust "anchor" that is less sensitive to outliers than a standard average.
- Distance-based Filtering: The server computes the Euclidean distance between each client's update and the median. Updates are flagged as benign only if their distance falls within a specific threshold ($\lambda$) relative to the median's magnitude.
- Clean Aggregation: The server averages only the updates that pass the filter. If no updates are deemed benign, the system defaults to the update closest to the median.
Experimental Evaluation
The researchers evaluated InferGuard across five benchmark datasets: MNIST, Fashion-MNIST, AT&T faces, SVHN, and GTSRB. The experiments simulated a non-IID environment where clients possess different data labels, reflecting realistic FL conditions. Effectiveness was measured using the Structural Similarity Index Measure (SSIM) and Learned Perceptual Image Patch Similarity (LPIPS), where lower scores indicate better defense (less recognizable reconstructed images).
Key Results and Discussion
InferGuard demonstrated superior protection compared to existing robust aggregation rules. A critical finding was the failure of popular methods like Multi-Krum. As shown in the comparison below, Multi-Krum frequently selects the malicious update (indicator = 1), whereas InferGuard successfully identifies and rejects it (indicator = 0) after the initial training rounds.
On the MNIST dataset, the SSIM of reconstructed images dropped to 0.22 with InferGuard, significantly lower than the 0.40+ seen in baseline methods. Qualitatively, while other defenses allowed the attacker to reconstruct clear digits, InferGuard reduced the attacker's output to unrecognizable noise.
Importantly, InferGuard maintains high global model accuracy, comparable to the standard FedAvg method, avoiding the utility trade-offs common in privacy-preserving techniques like Differential Privacy.
Conclusion
InferGuard provides a robust and practical defense against GAN-based data reconstruction attacks in Federated Learning. By treating privacy-leaking updates as a form of model poisoning, it effectively filters malicious contributions. Future work will focus on providing formal mathematical proofs for these filtering rules and developing automated methods for hyperparameter tuning to enhance "plug-and-play" deployment in large-scale models.
Summary updated