TL;DR
This paper introduces PPFPL, a novel framework for privacy-preserving federated learning (PPFL) that tackles the challenge of data poisoning attacks in cross-silo settings with non-independent and identically distributed (Non-IID) data. PPFPL leverages prototype learning and a secure two-server aggregation protocol to achieve robust model training while safeguarding client privacy.
What Problem Does This Paper Address?
Federated Learning (FL) allows multiple clients to collaboratively train a shared model without sharing their raw data, preserving privacy. However, in "cross-silo" FL, where clients are typically large organizations, this paradigm is vulnerable to data poisoning attacks. Malicious clients can manipulate their local data to submit poisoned model updates, degrading the performance of the global model. Existing privacy-preserving FL (PPFL) solutions, while offering strong privacy guarantees, often struggle to defend against these attacks, especially when client data is Non-IID. Non-IID data itself introduces variations in model updates, making it difficult to distinguish between genuine variations and malicious manipulations. This paper aims to bridge this gap by proposing a PPFL framework that is robust to data poisoning attacks even under Non-IID data distributions, while maintaining strong privacy guarantees.
Methods
PPFPL introduces two key innovations:
- Prototype Learning for Local Updates: Instead of submitting raw model updates (gradients), clients submit "prototypes." Prototypes are class-level feature representations that are less sensitive to data distribution shifts compared to gradients. This helps mitigate the impact of poisoned data distributions. A novel local optimization function is proposed that encourages local prototypes to align with global prototypes, using cosine similarity to constrain the influence of malicious clients.
- Secure Two-Server Aggregation Protocol: To aggregate these prototypes securely and robustly, PPFPL employs a two-server architecture (Aggregator and Verifier) that utilizes Homomorphic Encryption (HE) and Secure Multi-Party Computation (SMC). This protocol verifies the normalization of encrypted prototypes and then performs a Byzantine-robust aggregation, effectively filtering out malicious prototypes.
Datasets
The paper evaluates PPFPL on three standard public datasets:
- MNIST
- Fashion-MNIST (FMNIST)
- CIFAR10
Key Insights and Ideas
- Prototypes as a Defense: The core idea that prototypes, being class-level feature representations, are inherently more robust to data distribution shifts than gradients is a key insight. This makes them a promising avenue for defending against poisoning attacks in Non-IID settings.
- Two-Server Architecture for Robustness: The use of a non-colluding two-server architecture for aggregation is crucial. It allows for verification and secure computation, enabling the filtering of malicious updates without compromising privacy.
- Cosine Similarity for Malicious Client Constraint: Incorporating cosine similarity in the local optimization function is a clever way to constrain the influence of malicious clients by penalizing deviations from the trusted direction.
- Empirical Validation of Robustness: The experimental results consistently show that PPFPL maintains high accuracy even under significant data poisoning attacks and Non-IID conditions, often outperforming existing robust and privacy-preserving FL schemes.
Key Results
- High Accuracy under Attacks: PPFPL achieves accuracy comparable to or even exceeding the performance of non-attacked models in some cases, demonstrating its robustness against feature and label attacks on MNIST, FMNIST, and CIFAR10 under various Non-IID settings.
- Superior to Baselines: PPFPL consistently outperforms FedAvg (without attacks) and other robust FL schemes (e.g., ShieldFL, PBFL, FedDMC, AntidoteFL) in poisoned Non-IID scenarios.
- Efficient Parameter Submission: PPFPL significantly reduces the number of parameters submitted by clients compared to gradient-based methods, leading to lower communication and computational overhead.
- Stable Convergence: The loss curves show that PPFPL converges smoothly and is less affected by data poisoning attacks compared to some other methods.
Limitations and Future Work
- Two-Server Interaction Latency: The paper acknowledges that frequent interactions between the two servers can increase system failures due to communication latency.
- Dataset Complexity: Existing datasets may not fully capture the complexity of real-world cross-silo scenarios, potentially limiting the evaluation of benign client prototypes under varying data quality.
- Future Work: The authors propose two main directions:
- Designing more efficient two-server interaction protocols while maintaining privacy and security.
- Developing benchmark datasets that better represent cross-silo characteristics to evaluate FL system performance more accurately.
Summary updated