The Subsample Size Trade-Off in Robust Ensembles

14 concepts11 papers

When defending machine learning models against adversarial perturbations or data poisoning, a natural strategy is to avoid putting all your eggs in one basket. Instead of training a single monolithic model, systems can train an ensemble of sub-models on random subsamples of data, or randomly sample parts of the input at inference time, and aggregate their predictions via majority voting.

The core intuition behind this approach is isolation. If an attacker only controls a small number of poisoned points or adversarial features, those malicious elements will only appear in a fraction of the subsamples. The remaining uncorrupted sub-models will outvote the compromised ones, preserving the correct prediction.

However, designing these defenses reveals a sharp, unavoidable tension between clean accuracy and certified security. This tension is governed by a single tuning knob: the size of the subsample.

The Fundamental Tension: Fragmentation vs. Contamination

To understand why this trade-off exists, we can look at the two extremes of the subsample size knob:

  1. Small Subsamples (High Security, Low Accuracy): If each sub-model is trained on a very small subset of data, the probability that any given subset contains a poisoned point is remarkably low. This isolates the attack and allows the system to guarantee safety against a larger number of total adversarial interventions. But there is a catch: deep learning models are data-hungry. As shown by Nicola Bena et al., partitioning data too finely reduces the cardinality of individual training sets, causing severe data fragmentation that degrades the baseline performance of each sub-model.
  2. Large Subsamples (Low Security, High Accuracy): If we increase the subsample size, individual sub-models get enough data to learn rich, accurate decision boundaries on clean inputs. However, larger subsamples act like a wider net—they are much more likely to catch the attacker's poisoned points. If a poisoned point is included in too many sub-models, the adversarial influence spreads, collapsing the ensemble's collective defense.

Cross-Domain Manifestations

This fundamental trade-off is not unique to a single type of attack; it emerges across entirely different domains of machine learning security.

Data Poisoning and Bagging

In centralized learning, Jinyuan Jia et al. proved that classical bootstrap aggregating (bagging) possesses an intrinsic certified robustness against data poisoning. They formalize how the training subsample size acts as a direct mathematical lever balancing clean utility against poisoning sensitivity. Similarly, Marco Anisetti et al. leverage this property to protect Random Forests from label-flipping attacks, showing that optimizing this trade-off allows for resource-efficient defenses.

Federated Learning

In distributed environments where malicious clients can upload poisoned updates, Xiaoyu Cao et al. proposed training global models on random subsets of clients. They demonstrate that larger client subsets improve clean accuracy but simultaneously degrade the system's certified security level. In a follow-up work on provably secure federated learning, Xiaoyu Cao et al. highlight that partitioning clients into too many disjoint groups causes individual models to suffer from insufficient data, creating a strict upper bound on practical security.

3D Point Cloud Classification

The trade-off even applies to test-time adversarial defenses. For instance, Hongbin Liu et al. introduced a provably robust framework for 3D point clouds by randomly subsampling points at inference time. While smaller point subsets effectively isolate adversarial geometric distortions, they also strip away the fine-grained structural details the model needs to correctly identify complex 3D objects.

The Limits of Pure Scaling

If shrinking the subsample size hurts accuracy, can we fix it simply by training more sub-models? Intuitively, generating thousands of tiny, diverse models might seem like a way to recover the lost performance.

However, research shows that ensemble defenses face severe diminishing returns to scale. Simply increasing the cardinality of an ensemble does not monotonically increase its robustness.

As Hongwei Zhang et al. observe, while adding more models increases computational costs linearly, the reduction in the shared adversarial subspace quickly plateaus. Jie Shi et al. reinforce this by demonstrating that sheer ensemble size does not guarantee better defense; rather, the underlying transformation or diversity mechanism matters more.

This saturation effect is a recurring theme across various ensemble architectures:

  • Sailik Sengupta et al. find diminishing security benefits once a moving target defense reaches a certain architectural diversity.
  • the authors of Jiang Liu et al. note that finding the optimal cohort size to balance computational overhead with resilience remains an open challenge.
  • Even when scaling defensive objectives or adapters rather than distinct models, Kangtao Lv et al. show that over-optimizing or over-scaling multi-method combinations can cause decision boundaries to overfit, paradoxically degrading overall robustness.

Ultimately, robust ensemble design is not a free lunch. Security cannot be achieved by infinitely scaling model counts or blindly partitioning datasets. Instead, practitioners must carefully navigate the Pareto frontier defined by the subsample size—balancing the statistical necessity of large datasets against the security imperative of strict data isolation.

Key Foundational Work

Where Does the Robustness Come from?: A Study of the Transformation-based Ensemble Defence (2020) — contributed 2 concept elements to this direction.
MTDeep: Boosting the Security of Deep Neural Nets Against Adversarial Attacks with Moving Target Defense (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_subsample_security_trade_off_8661f4d1,
  title = {subsample-security trade-off},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/subsample-security-trade-off/rdr_8661f4d11f834c5e8474ff196223cc9f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_subsample_security_trade_off_8661f4d1_v1,
  title = {subsample-security trade-off},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/subsample-security-trade-off/rdr_8661f4d11f834c5e8474ff196223cc9f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake