Introduction
Federated Learning (FL) has emerged as a powerful paradigm for training machine learning models on distributed user data without requiring raw data to leave local devices. However, this decentralized approach faces a fundamental "privacy vs. robustness" conflict. To ensure privacy, individual model updates should be obscured from the central server. Conversely, to ensure robustness, the server must inspect these updates to identify and filter out "Byzantine" (malicious) users who might attempt to poison the global model or inject backdoors.
Existing solutions typically fall into two traps: they either employ weak security measures (such as simple norm-checks) that advanced attackers can easily circumvent, or they utilize complex robust aggregation algorithms (like multi-Krum) that inadvertently leak sensitive information, such as the distances between users' private updates. PriRoAgg addresses this gap by introducing a framework that achieves advanced robustness while maintaining a strict standard of privacy.
The Concept of Aggregated Privacy
The core innovation of PriRoAgg is the formalization of Aggregated Privacy. This security notion is based on the philosophy of "minimum leakage": the server should only learn the absolute minimum statistical information necessary to perform robust aggregation. Instead of viewing individual updates, the server only sees the sum of specific functions of user data (e.g., the sum of the signs of gradients). This allows for outlier detection and robust filtering without compromising the confidentiality of any single user's contribution.
The PriRoAgg Framework
PriRoAgg utilizes three primary technical pillars to execute robust aggregation while satisfying aggregated privacy:
- Lagrange Coded Computing (LCC): A specialized form of secret sharing that allows users to split their updates into "shares." This enables the server to perform linear computations on the encrypted pieces. LCC is significantly more efficient than standard Homomorphic Encryption and allows for the partitioning of model updates into $K$ pieces, reducing communication overhead by a factor of $K$.
- Distributed Zero-Knowledge Proofs (SNIP): To prevent malicious users from submitting fraudulent secret shares, PriRoAgg employs an optimized version of SNIP. This allows users to prove they followed protocol rules (e.g., "I correctly calculated the sign of my gradient") without revealing the gradient itself. The authors optimized these proofs by exploiting the "repeated patterns" in model updates, where the same mathematical operations are applied across every coordinate.
- Protocol Instantiations: The framework is instantiated into two specific protocols:
- PriRoAgg-RLR: Utilizes sign-based information to defend against sophisticated backdoor attacks.
- PriRoAgg-RFA: Employs a geometric median approach to defend against model poisoning.
Experimental Evaluation
The researchers evaluated PriRoAgg using standard computer vision benchmarks, Fashion-MNIST and CIFAR-10, against a variety of adversarial threats, including Gaussian noise, update scaling, and "Min-Max" poisoning.
Robustness Performance
In a "Min-Max" poisoning attack on CIFAR-10, PriRoAgg demonstrated superior resilience, maintaining 79.2% accuracy. In contrast, existing secure protocols like RoFL and EIFFeL collapsed under the same conditions, with accuracies dropping as low as 10-18%. Furthermore, against the "Trojan A" backdoor attack on Fashion-MNIST, PriRoAgg achieved an Attack Success Rate (ASR) of 0.0%, effectively neutralizing the threat while preserving the accuracy of the main task.
Efficiency and Scalability
As the number of users ($N$) and the dimensionality of the model ($d$) increase, PriRoAgg scales more favorably than its predecessors. By leveraging LCC and optimized zero-knowledge proofs, the framework significantly reduces the per-iteration computational and communication overhead for large-scale deep learning models compared to baselines like EIFFeL.
Conclusion and Future Work
PriRoAgg provides a robust and private framework for Federated Learning by revealing only the necessary aggregated statistics for model integrity. While the current work assumes a "semi-honest" server—one that follows the protocol but attempts to learn extra information—it sets a new standard for balancing the conflicting requirements of FL. Future research may explore extending this framework to non-colluding multi-server scenarios to further harden privacy against active, malicious servers and expanding the library of supported robust aggregation algorithms.
Summary updated