Securing Collaborative Learning Against Malicious Participants

931 concepts773 papers14 proposals

In the design of privacy-preserving machine learning (PPML), researchers must define what an adversary is allowed to do. The vast majority of current frameworks—ranging from secure inference to federated training—rely on the semi-honest (or "honest-but-curious") adversary model. Under this assumption, all participants are expected to follow the protocol's mathematical steps exactly, but they are allowed to inspect the messages they receive to try and infer private information.

While this assumption enables high-performance systems, it creates a significant security gap. As noted by Lou et al., semi-honest systems do not necessarily provide protection against malicious adversaries who might actively manipulate the protocol to subvert privacy or accuracy.

The Vulnerability of the Semi-Honest Assumption

The semi-honest model is popular because it minimizes the computational "tax" of cryptography. By assuming participants won't lie about their inputs or skip steps, researchers can use efficient primitives like additive secret sharing or optimized polynomial approximations. For example, Lou et al. and Lin demonstrate how semi-honest frameworks can achieve massive speedups in inference and classification by focusing on optimization rather than active verification.

However, this trust is fragile. In real-world decentralized environments, a single participant might deviate from the rules to gain an advantage. The authors of MP-FedXGB point out that the honest-but-curious assumption leaves systems susceptible to active subversion, such as injecting poisoned data to compromise the resulting model. Similarly, Jiang et al. highlight that secure aggregation protocols are often vulnerable to active poisoning attacks where malicious clients submit corrupted gradients to sabotage global training.

The Malicious Adversary Challenge

A malicious security model assumes the worst: participants may provide fraudulent data, collude with others, or intentionally abort the protocol to leak information. Transitioning from semi-honest to malicious security is not a simple upgrade; it typically requires robust verification mechanisms to prove that every step of the computation was performed correctly.

As Zhang et al. argue, this transition remains an open challenge because the required verification mechanisms introduce high computational overhead. This "efficiency gap" is a recurring theme:

  • Lu et al. note that extending high-performance secure inference to handle malicious actors introduces substantial communication overhead.
  • Shamsabadi et al. observe that protecting against active cheating often hinders practical performance in wide-area networks.
  • Acar et al. emphasize that without these robust defenses, distributed systems remain fundamentally exposed to data poisoning.

Bridging the Gap

Recent research focuses on finding middle ground—achieving "malicious-grade" security without the prohibitive costs of traditional zero-knowledge proofs or heavy verification. Zheng et al. proposed Helen, a framework that achieves robust security even when a majority of participants are actively malicious, ensuring that private data remains confidential and the model remains accurate despite intentional subversion.

Other approaches attempt to mitigate specific malicious behaviors, such as data poisoning, without full cryptographic verification. For instance, Byrd explores combining Differential Privacy with Secure Multi-Party Computation to protect against colluding malicious participants in financial fraud detection.

Ultimately, the field is moving toward a realization that the semi-honest model, while a useful starting point for efficiency, is insufficient for high-stakes applications like healthcare (Chase et al.) or genomics (Cho et al.). The core intuition driving current research is that a truly secure system must not only keep secrets but also enforce the integrity of the collaborative process itself.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 756 papers
  2. Version 1 · 2026-10-04 · 675 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_semi_honest_adversary_model_658b1955,
  title = {semi-honest adversary model},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semi-honest-adversary-model/rdr_658b1955a6bc4224ac2217b38e6a3757},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_semi_honest_adversary_model_658b1955_v2,
  title = {semi-honest adversary model},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semi-honest-adversary-model/rdr_658b1955a6bc4224ac2217b38e6a3757/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake