In decentralized machine learning, a fundamental tension exists between securing a model against malicious attacks and ensuring it treats all legitimate users fairly. This conflict arises because, from the perspective of a central aggregator, honest diversity is statistically indistinguishable from adversarial malice.
Consider a heterogeneous (non-IID) federated learning environment where different clients possess distinct data distributions. A minority group might contribute updates that deviate sharply from the majority consensus. Crucially, an attacker attempting to poison the model or inject a backdoor will also submit updates that deviate from the consensus. As highlighted by Binghui Zhang et al. in FedTilt: Towards Multi-Level Fairness-Preserving and Robust Federated Learning, legitimate but rare data points often share similar statistical signatures with corrupted data, forcing a zero-sum trade-off between security and inclusivity.
The Defensive Filter: How Robustness Silences the Tail
To defend against Byzantine actors who send arbitrary or poisoned gradients, traditional distributed learning relies on robust aggregation rules. These rules—such as trimmed means, coordinate-wise medians, or clustering-based methods—operate by identifying and filtering out statistical outliers.
However, as Hongyi Wang et al. demonstrate in Attack of the Tails: Yes, You Really Can Backdoor Federated Learning, this outlier-filtering mechanism creates a severe exclusion bias. When a defense rejects updates that look anomalous to protect against edge-case backdoors, it simultaneously silences honest users who happen to own rare or edge-case data. El Mahdi El Mhamdi et al. prove in Collaborative Learning in the Jungle (Decentralized, Byzantine, Heterogeneous, Asynchronous and Nonconvex Learning) that statistical outlier detection used for Byzantine resilience inherently discards honest nodes representing minority viewpoints.
Majority Clients Minority Clients Malicious Actors
(Standard Gradients) (Unique Gradients) (Poisoned Gradients)
│ │ │
▼ ▼ ▼
┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ In-lier │ │ Outlier │ │ Outlier │
└──────┬───────┘ └──────┬───────┘ └──────┬───────┘
│ │ │
▼ └───────────┬──────────────┘
┌──────────────────────────────────────────────▼──────────────────────┐
│ Byzantine-Robust Aggregator (Outlier Filter) │
├──────────────────────────────────────────────┬──────────────────────┤
│ Kept │ Discarded │
│ (Biased toward Majority) │ (Silences Minority) │
└──────────────────────────────────────────────┴──────────────────────┘
This phenomenon is further analyzed by Wenke Huang et al. in Federated Learning for Generalization, Robustness, Fairness: A Survey and Benchmark, who note that prioritizing system robustness systematically undermines demographic fairness. Similarly, Jun Wu et al. show in BOBA: Byzantine-Robust Federated Learning with Label Skewness that robust aggregators frequently misidentify legitimate updates from minority classes as malicious attacks, biasing the global model toward the majority distribution.
Even when defenses incorporate privacy or quality checks, the trade-off persists:
- Farinaz Koushanfar et al. note in You Can Have Your Cake and Eat It Too: Ensuring Practical Robustness and Privacy in Federated Learning that strict robustness checks inadvertently filter out legitimate updates from diverse or edge-case data sources.
- Qi Li et al. demonstrate in martFL: Enabling Utility-Driven Data Marketplace with a Robust and Verifiable Federated Learning Architecture that if an aggregator relies on a biased reference dataset to filter updates, it will exclude valuable, non-conforming data from honest providers, limiting the model's overall fairness.
- Edith C. H. Ngai et al. show in Auto-weighted Robust Federated Learning with Corrupted Data Sources that filtering participants based on performance metrics incorrectly penalizes benign participants with legitimately difficult data distributions.
The Vulnerability of Inclusion: How Fairness Invites Exploitation
If filtering outliers harms fairness, the intuitive alternative is to explicitly protect the "tail" of the distribution. Fairness frameworks often achieve this by minimizing the worst-case risk across groups or optimizing for relative performance gains across clients.
However, optimizing for the worst-performing samples inherently increases a model's vulnerability to data corruption and manipulation. As Runtian Zhai et al. observe in Boosted CVaR Classification, specifically optimizing for the hardest samples forces a model to over-index on outliers and noisy labels.
In a distributed setting, this creates a dangerous vulnerability. A malicious actor can intentionally submit highly unrepresentative or poisoned data that yields a high loss. A fairness-oriented objective will interpret this high loss as a sign of an underrepresented, struggling client and will eagerly amplify the malicious update to ensure egalitarian performance. As Guojun Zhang et al. point out in Proportional Fairness in Federated Learning, without robust weighting mechanisms, fairness-oriented algorithms are highly vulnerable to being hijacked by participants with poor quality or unrepresentative data.
The mathematical limits of this tension are tight. Yuhao Yi et al. show in Near-Optimal Resilient Aggregation Rules for Distributed Learning Using 1-Center and 1-Mean Clustering with Outliers that increasing an algorithm's ability to ignore distant outliers simultaneously increases its vulnerability to subtle, biased updates hidden within the honest data distribution. Furthermore, Jiawen Weng et al. note in FedQP: Towards Accurate Federated Learning using Quadratic Programming Guided Mutation that while constraining updates toward a global gradient improves convergence, its impact on system fairness and vulnerability to security threats in real-world physical deployments remains an open question.
Breaking the Deadlock: Mechanisms for Joint Robustness and Fairness
To escape this zero-sum game, recent research focuses on architectures and aggregation rules that can distinguish genuine data diversity from malicious poisoning.
Decoupling via Personalization
One powerful approach is to move away from a single, rigid global model. In Ditto: Fair and Robust Federated Learning Through Personalization, Tian Li et al. propose a multi-task learning framework where clients learn personalized local models regularized toward a global model.
$$\min_{v_k} h_k(v_k) + \frac{\lambda}{2} \|v_k - w^*\|^2$$
By decoupling local performance from a single global objective, the global model ($w^*$) does not need to absorb extreme minority updates to be fair, nor does it need to aggressively filter them to be robust. Legitimate but rare data distributions are accommodated locally ($v_k$) without making the entire network vulnerable to malicious updates that mimic outliers.
Soft Penalization and Joint Optimization
Instead of binary filtering (keeping or discarding an update), recent methods use soft penalization or joint optimization. In Tolerating Outliers: Gradient-Based Penalties for Byzantine Robustness and Inclusion, Latifa Errami et al. introduce a linear scalarization technique that penalizes rather than discards suspected gradients. This allows the model to mitigate large adversarial shocks while still extracting useful learning signals from honest but unusual outlier data.
Similarly, Guanqiang Zhou et al. develop the H-nobs: Achieving Certified Fairness and Robustness in Distributed Learning on Heterogeneous Datasets framework. They combine fairness-weighted loss functions with a Norm-Based Screening (NBS) aggregation rule. This combination ensures that the system remains resilient to Byzantine nodes while actively improving the model's performance on the most challenging or underrepresented data partitions, providing the first convergence guarantees for systems that are simultaneously fair and robust.
Ordinal and Rank-Based Aggregation
Another paradigm shifts the focus from absolute gradient values to ordinal preferences. In Fair Federated Learning via the Proportional Veto Core, Bhaskar Ray Chaudhury et al. adapt concepts from social choice theory to create the Rank-Core-Fed algorithm. By aggregating how clients rank candidate models rather than averaging their raw updates, the system provides inherent protection against model hijacking by a small number of outliers, ensuring that even highly heterogeneous agents receive a model they rank highly without compromising the preferences of the majority.
Key Foundational Work
Attack of the Tails: Yes, You Really Can Backdoor Federated Learning (2020) — contributed 4 concept elements to this direction.
Ditto: Fair and Robust Federated Learning Through Personalization (2021) — contributed 3 concept elements to this direction.
Collaborative Learning in the Jungle (Decentralized, Byzantine, Heterogeneous, Asynchronous and Nonconvex Learning) (2021) — contributed 2 concept elements to this direction.