Skip to content
Crypto.world

Research

Peer-Relative Trust-Guided Update Screening for Robust Semi-Decentralized Federated Learning under Label-Flipping Attacks

In federated learning, clients train locally and send model updates to a server, and a few malicious clients can degrade the shared model by tampering with their updates — here, by persistently…

Peer-Relative Trust-Guided Update Screening for Robust Semi-Decentralized Federated Learning under Label-Flipping Attacks
arXiv
2610.25714
Published
2026-10-19
Authors
Ajay Kumar Shrestha, Sandhya Joshi

AI summary

Overview

Research area: Security and robustness of federated learning — specifically trust and reputation mechanisms in semi-decentralized federated learning (SDFL) under label-flipping attacks, with a view toward blockchain-based audit and incentive layers.

Technical level: Intermediate. The abstract assumes familiarity with federated learning, non-IID data partitioning, and standard attack/defense terminology, but the core idea (screening client updates using relative trust) is explained in accessible terms.

Scope: The paper turns previously proposed trust signals into an operational update-screening mechanism and measures, on an MNIST benchmark, how well it separates and rejects malicious clients while keeping honest ones under two levels of data heterogeneity.

What This Paper Is About

In federated learning, clients train locally and send model updates to a server, and a few malicious clients can degrade the shared model by tampering with their updates — here, by persistently flipping labels. Trust and reputation scores can help, but simply filtering out updates that look bad is risky: when clients' data is unevenly distributed (non-IID), an honest client's update can temporarily hurt performance on a shared validation set and be mistaken for an attack. The paper's goal is to screen updates using peer-relative trust signals — comparing a client against its peers rather than against a single absolute threshold — so that malicious updates are rejected without discarding too many honest ones.

Key Contributions

  1. Operationalizes prior trust signals. Building on the authors' earlier trust-based semi-decentralized FL framework, the paper turns previously conceptual trust and reputation signals into concrete, usable screening criteria.
  2. Introduces peer-relative update screening. The proposed method combines a robust validation-utility score with two additional update-quality signals: the direction of the update and its norm (magnitude).
  3. Quantifies the utility–selectivity trade-off. The evaluation explicitly measures how much accuracy is retained versus how many malicious updates are blocked, and contrasts screening against trust weighting alone — separating "influence mitigation" (reducing the weight of bad clients) from "explicit containment" (actually blocking their updates).
  4. Provides an off-chain governance component for SDFL. The screening procedure is positioned as an empirically validated off-chain core that can supply blockchain-based audit and incentive layers with trustworthy decisions.

Main Findings

  • Strong accuracy under moderate heterogeneity: With Dirichlet heterogeneity of 0.5, the method reaches 98.37% final accuracy across 50 MNIST clients at 100 communication rounds, with 20% persistent label-flip attackers.
  • High rejection with good benign retention (moderate case): It rejects 95.82% of malicious updates while still accepting 90.30% of benign updates.
  • Trust scores fully separate attackers (moderate case): The final benign and malicious trust distributions are completely separated, with a Kolmogorov–Smirnov statistic of KS = 1.00.
  • Degradation under severe heterogeneity: At heterogeneity 0.1, accuracy falls to 95.25%, malicious rejection drops to 88.55%, and benign retention drops to 76.90%, with trust separation slightly weaker at KS = 0.96.
  • Screening costs some utility relative to weighting alone: Trust weighting by itself achieves slightly higher utility in the most heterogeneous condition, but it accepts all malicious submissions — meaning it mitigates influence without containing attackers.
  • The central trade-off is quantified: Screening buys explicit containment at the price of some benign updates being rejected, and the paper frames this utility-versus-selectivity trade-off as its main empirical message.

Methodology in Plain English

The authors set up a simulated semi-decentralized federated learning system with 50 clients training on MNIST, one-fifth of which are persistent label-flip attackers — clients that deliberately corrupt the labels in their local data so their updates pull the global model in the wrong direction. They vary how unevenly the data is spread across clients using a Dirichlet partition at two levels: a moderate setting (0.5) and a severe setting (0.1), running 100 communication rounds and repeating each experiment over five paired random seeds so results are compared under matched conditions.

Rather than judging an update in isolation, the proposed screening judges it relative to its peers. Each incoming update is scored along three lines: how useful it proves on a robust validation measure, which direction it points in update space, and how large it is (its norm). Updates that look anomalous on these peer-relative criteria are screened out; the rest are kept. The authors then track three things over training: final model accuracy, what fraction of malicious updates were blocked, what fraction of benign updates survived the filter, and how far apart the trust scores of benign and malicious clients ended up.

The abstract does not report the exact scoring formula, thresholds, the mechanism by which peers are compared, or any comparison against standard aggregation baselines beyond the trust-weighting-only variant.

Why This Matters

Impact on research: The paper reframes an often-conflated distinction in federated learning security — reducing a bad client's influence versus actively containing it — and reports both sides of the trade-off rather than only accuracy. Its finding that a simple trust-weighting approach can match or slightly beat screening on utility while letting every malicious update through is a useful caution for work that reports accuracy alone.

Real-world applications:

  • Cross-institution model training where participants cannot share raw data but may have incentives to poison the shared model, such as hospitals or banks collaborating on a joint model.
  • Semi-decentralized or consortium federated systems that need an auditable record of which participants' updates were accepted or rejected.
  • Blockchain-anchored incentive schemes, where the screening decision determines whether a client is rewarded or penalized — the paper explicitly targets this use.
  • Deployments with naturally heterogeneous client data (mobile keyboards, edge devices, regional user populations), where honest updates are most likely to be wrongly flagged.

Industry relevance: Organizations running federated learning across untrusted or semi-trusted parties need a defensible, measurable filter before updates influence a production model. A component that produces explicit accept/reject decisions and trust scores is directly pluggable into audit logs and token- or reputation-based reward systems.

Future Directions

  • Improve benign retention under severe heterogeneity. Retaining only 76.90% of honest updates at heterogeneity 0.1 is the clearest weak point; better peer-relative criteria or adaptive thresholds could recover these.
  • Bridge the off-chain/on-chain gap. The abstract presents the screening core as something that can feed blockchain audit and incentive layers; actually implementing and evaluating that integration — including how trust scores are recorded and acted upon — remains open.
  • Test beyond the reported setting. The evaluation is confined to MNIST, label-flip attacks, a 20% attacker fraction, and 100 rounds; other datasets, attack types, attacker proportions, and longer training horizons are untested here.
  • Compare against a wider set of defenses. The abstract's only stated comparison is trust weighting alone; benchmarking against other robust aggregation or filtering schemes would clarify where the method sits.

Target Audience

Researchers and graduate students working on federated learning security, robust aggregation, and trust/reputation design; engineers building federated or semi-decentralized learning systems across mutually distrusting parties; and blockchain researchers interested in how off-chain governance signals can be made reliable enough to drive on-chain audit and incentive mechanisms. Readers without federated learning background will need to acquire the basics of non-IID data partitioning and model-update attacks first.

Authors’ abstract

Trust and reputation mechanisms can reduce the influence of malicious clients in federated learning (FL), but strict update filtering is difficult under non-independent and identically distributed (non-IID) data because honest local updates may temporarily reduce performance on a shared validation proxy. Building on our earlier trust-based semi-decentralized FL (SDFL) framework, this paper operationalizes the trust signals and introduces peer-relative update screening that…

The full abstract on arXiv
Read the original paper