Introduction
Federated Learning (FL) enables collaborative model training without direct data sharing, offering significant privacy advantages. However, its decentralized nature introduces vulnerabilities, particularly to backdoor attacks. In these attacks, malicious participants submit poisoned model updates that embed specific triggers into the global model, causing it to misclassify targeted inputs while maintaining high accuracy on standard data.
Existing defenses often struggle in non-IID (not identically distributed) environments. They frequently fail in two ways:
- Insider Threats: Subtle malicious updates can bypass initial anomaly detection filters.
- Benign Exclusion: Legitimate clients with unique data distributions are often misclassified as malicious, degrading model performance and fairness.
The FedDLAD Framework
To address these challenges, the authors propose FedDLAD (Federated Learning Dual-Layer Anomaly Detection), a four-stage defense mechanism designed for robustness and fairness.
1. COF-Based Anomaly Detection & 2. IQR-Based Secondary Detection
Initial stages of FedDLAD: COF-based structural analysis and IQR-based dimension-wise filtering with sign flipping. Source, p. 3
The server first utilizes the Connectivity-Based Outlier Factor (COF) to evaluate the structural connectivity of incoming model parameters. This provides an initial categorization of clients into "trusted" and "untrusted" groups. To identify malicious "insiders" who may have bypassed the first filter, FedDLAD employs an Interquartile Range (IQR) filter. This module performs a dimension-wise inspection of updates. Instead of discarding anomalous values, it flips their sign, effectively neutralizing malicious influence while retaining the update's structural information.
3. The Pardon Module & 4. Final Aggregation
Final stages of FedDLAD: Re-evaluating untrusted clients via the Pardon Module and performing weighted aggregation for the global model update. Source, p. 3
To mitigate benign exclusion, the framework incorporates a "pardon" mechanism. A "reference update" is synthesized from the trusted group. The server then calculates the cosine similarity between this reference and updates in the untrusted group. Clients exhibiting high similarity are "pardoned" and reintegrated. Finally, the global model is updated through a weighted aggregation of the reference update and the pardoned updates, ensuring that contributions from diverse but legitimate clients are preserved.
Experimental Evaluation
The framework was tested on four benchmarks: MNIST, FashionMNIST, SVHN, and CIFAR-10. Data heterogeneity was simulated using a Dirichlet distribution among 50 clients, with a 20% malicious client ratio.
Key Results
FedDLAD demonstrates superior performance compared to state-of-the-art defenses like Krum, RLR, and FLTrust:
- Backdoor Mitigation: On the non-IID SVHN dataset, FedDLAD reduced the Attack Success Rate (ASR) to 4.4%, a sharp contrast to the 81% seen with standard Federated Averaging (FedAvg).
- Accuracy Maintenance: It maintained a Natural Accuracy (ACC) of 89.5% in the SVHN scenario, outperforming competitors that often sacrifice accuracy for security.
- Scalability: The framework remained robust even when the proportion of malicious clients reached 50%.
Discussion and Insights
A key innovation in FedDLAD is the use of sign flipping for anomalous dimensions. This active defense mechanism uses the attacker's own gradient direction to pull the model back toward a benign state. Furthermore, the Pardon Module proves essential for non-IID settings, showing that security does not have to come at the cost of excluding legitimate, diverse data. Ablation studies highlight the specialized roles of each layer: the COF module excels in stable (IID) environments, while the IQR module is critical for resilience in highly diverse (non-IID) scenarios.
Conclusion and Future Work
FedDLAD provides a robust, dual-layer defense against backdoor attacks in Federated Learning, particularly in challenging non-IID environments. While highly effective, the authors note that extreme data heterogeneity still poses challenges for reference update generation. Future research will focus on refining efficiency and addressing adaptive attacks where adversaries may attempt to circumvent the dual-layer defense.
Summary updated