Reconciling Adversarial Robustness and Fairness in Distributed Learning

This direction is split into other directions: robustness-fairness trade-off. Its last version stays readable and citable here.

Version 1 of 126 papers45 conceptsfrozen 2026-10-04

In distributed and federated learning, a central server aggregates model updates from many participants to build a shared global model. This process faces two fundamental challenges: robustness (protecting the model from malicious "Byzantine" actors who send poisonous updates) and fairness (ensuring the model performs well for all participants, including those with rare or minority data).

While these goals seem complementary, recent research reveals they are often in direct conflict. The mechanism used to achieve one frequently undermines the other, creating a "reliable gap" in real-world deployments.

The Outlier Dilemma

The core of the conflict lies in how we treat "outliers." In a typical robust aggregation scheme, the server identifies and filters out updates that are statistically distant from the average. The intuition is simple: if an update looks nothing like the others, it is likely malicious.

However, as Michael-Andrei Panaitescu-Liess and Tudor Dumitras argue in their work Like Oil and Water, legitimate minority groups also produce updates that look like outliers. Because these groups have data distributions that differ from the majority, their honest gradients often have high norms or unusual directions.

When a system applies strict poisoning defenses, it inadvertently "silences" these minority participants. This leads to a significant drop in accuracy for underrepresented groups, effectively sacrificing fairness for the sake of security. Conversely, El Mahdi El Mhamdi and Sébastien Rouault note that inclusive systems—those designed to represent diverse viewpoints—are often more vulnerable to "jungle-like" environments where malicious nodes can hide behind the mask of legitimate heterogeneity.

The Attack of the Tails

This tension creates a security vulnerability known as the "edge-case backdoor." Hongyi Wang and Kartik Sreenivasan demonstrated in Attack of the Tails that attackers can craft malicious updates that specifically target the "tails" of the data distribution.

Because these attacks mimic the statistical profile of rare but legitimate data, they are nearly impossible to detect using standard outlier-based filtering. If the defender is too aggressive, they exclude honest minority users; if they are too permissive, they allow the backdoor to take hold. This suggests that robustness and fairness cannot be solved in isolation; they are two sides of the same distributional coin.

Bridging the Gap through Personalization and Screening

To resolve this tension, researchers are moving away from a "one-size-fits-all" global model.

Personalized Regularization

Tian Li and Virginia Smith introduced Ditto (Ditto: Fair and Robust Federated Learning Through Personalization), a framework that decouples local model performance from the global objective. In Ditto, participants learn a personalized local model that is regularized toward a robust global model. This allows the global model to remain secure by filtering outliers, while the local models retain the ability to adapt to the specific, rare nuances of the user's own data.

Norm-Based Screening

Another approach involves more sophisticated aggregation rules that provide theoretical guarantees for both properties. Guanqiang Zhou and Zhi Tian developed H-nobs (Achieving Certified Fairness and Robustness in Distributed Learning), which uses a technique called Norm-Based Screening (NBS). Instead of simply discarding high-norm updates, NBS uses fairness-promoting objectives to distinguish between a malicious poisoning attempt and a legitimate, high-loss gradient from a minority user.

Toward Reliable Federated Learning

The shift in the field, as highlighted by the survey work of Wenke Huang and Mang Ye, is toward a unified view of "reliability." A truly reliable system must account for the fact that in a heterogeneous world, the "average" is a poor representation of the whole.

Whether through rank-based mechanisms like Rank-Core-Fed (Bhaskar Ray Chaudhury et al.) or hybrid penalties that "tolerate" outliers rather than deleting them (Latifa Errami and El Houcine Bergou), the goal is to build models that are secure precisely because they are inclusive.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_reconciling_adversarial_robustness_and_f_eca2e677,
  title = {Reconciling Adversarial Robustness and Fairness in Distributed Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reconciling-adversarial-robustness-and-fairness-in-distributed-learning/rdr_eca2e67777e546df8c70c23f69d6cc8d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_reconciling_adversarial_robustness_and_f_eca2e677_v1,
  title = {Reconciling Adversarial Robustness and Fairness in Distributed Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reconciling-adversarial-robustness-and-fairness-in-distributed-learning/rdr_eca2e67777e546df8c70c23f69d6cc8d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}