acceptodds
Under review as a conference paper at ICLR 2027

ProtegoFed: Sample-Level Backdoor Defense for Federated Instruction Tuning with Interspersed Poisoned Data

Abstract

Federated instruction tuning (FIT) enables organizations to adapt large language models without sharing private instructions. Its training data, however, may be collected from untrusted users or third parties, allowing poisoned samples to be interspersed across otherwise benign clients. Existing federated backdoor defenses focus mainly on malicious clients and can be ineffective in this setting. We identify frequency-domain sample gradients as a signal for separating poisoned from clean instructions and introduce ProtegoFed, a one-time, three-stage defense that combines intra-client clustering, server-side secondary clustering of client centroids, and local revision using a global centroid. Across four free-style question answering datasets and four backdoor attacks, ProtegoFed identifies – of poisoned samples in the IID setting, reduces attack success rate to zero in all 16 dataset-attack combinations, and largely preserves main-task utility in those settings. Additional experiments cover nine models from 1B to 70B parameters, heterogeneous client distributions, clean-only data, design ablations, adaptive attacks, and integration with a client-level defense.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
FL, FIT, Backdoor, Identification, Defense, Poison

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_6whPOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors present ProtegoFed, which is a sample level backdoor defense for Federated Instruction Tuning with Interspersed Poisoned Data. Here the adversary is not a client, but a data provider so poisoned instruction records end up mixed with the clean ones inside benign clients. Existing FL defenses that work on targetting malicious clients do not apply here. The proposed defense works by running once before tuning happens and uses sample-wise gradients on lora-B of the last transformer layer. The authors evaluate this on 4 datasets and 4 attacks over a wide range of experimental settings. Overall I like the approach, but my main concern is the IID setting which I elaborate below.

Critical strengths
  • The threat model is well chosen because the record-level contamination inside the benign silos is a different problem from Byzantine clients.

  • The breadth of evaluation is wide, nine models, four attack families, and several ablations. The authors have covered a variety of experimental settings to ensure that the technique is generalizable.

  • I like how the centroids are computed in the frequency space, because the per-client embeddings are not mutually comparable. Overall the formulation and approach to the solution in the paper is solid.

Critical weaknesses
  • My main concern is the NIID setting. the NIID setting in FL almost always refers to the data distribution heterogeneity, and not the poison ratio heterogeneity. In that sense, the paper does not test the scenario where the clients hold different task or domain distributions.

  • The ASR being 0.00 across so many experiments seem odd. Is this for one seed. There should be multiple experiments and a variance should also be reported in the main paper.

  • All the main experiments in the paper are run in an IID cross-silo setting. This is another fundamental objection I have because in practice FL is never on IID data, and a defense will always perform better on an IID dataset. An attacker will always be weak in the IID setting. These are fundamental concepts in the FL community.

Questions for authors
  • How well does the method work in cross-device settings?

  • What are the seeds used for generating tables 2-4 and what are the standard deviations/variances for them?

  • Can you report the main results with non-iid distributions, and not just poison heterogeneity? This is my main concern.

Reviewer_hH5MOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose ProtegoFed, a method for removing poisoned examples before federated instruction tuning. Each organization clusters frequency-transformed per-example gradients and sends a cluster summary to the server. The server constructs a shared reference that organizations use to revise their local filtering decisions. The evaluation covers four QA datasets and four backdoor attacks, with additional experiments on different models, heterogeneous data distributions, and adaptive attacks.

Critical strengths

I liked that the paper evaluates four attack types on four QA datasets. Across the 16 main evaluation settings, ProtegoFed reports 0.00% attack success while largely maintaining clean-task performance. I also liked the ablation measuring the value of coordination across organizations: under NIID-3, poisoned-example recall improves from roughly 50% without global correction to nearly 100% with the full method.

Critical weaknesses

The contribution of the frequency transform is unclear. I would like a comparison using raw gradients, with and without normalization, while keeping the clustering and federated coordination unchanged. This would help establish what the DCT adds to the method.

A simple text-based filtering baseline is missing. The evaluated attacks introduce recurring phrases or responses into poisoned answers, including in the additional target-output experiments. A baseline based on the frequency of answer phrases would help establish the added value of gradient-based clustering and federated coordination. Including legitimate repeated responses would make this comparison more informative, since repetition alone does not imply poisoning.

The evaluation does not establish effectiveness when organizations perform different tasks. Although the authors test non-IID distributions within QA datasets, it remains unclear whether the shared clean reference preserves legitimate examples when organizations contribute different tasks, such as summarization, translation, and QA. Such task differences could affect gradient patterns independently of poisoning.

Questions for authors

How does ProtegoFed perform when the DCT is removed and the same pipeline uses raw or normalized gradients?

How does ProtegoFed compare with a simple repeated-phrase filtering baseline, particularly when clean examples also contain recurring responses or templates?

Have you evaluated a federation in which organizations contribute different task types? How would this affect clean-example retention, task performance, and attack success?

Reviewer_Me2tOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PROTEGOFED, which filters poisoned instruction samples before federated training through local gradient clustering, server clustering of client centroids, and local revision using a global centroid.

Critical strengths

1.The paper studies a practical threat in which poisoned samples are distributed across otherwise benign clients.

2.The overall pipeline is straightforward and does not require sharing raw instruction records.

3.The experiments cover multiple models, attack constructions, and heterogeneous data distributions, with strong results in the evaluated settings.

Critical weaknesses

1.The paper's contribution beyond existing work is insufficiently clear, GraCeFul already uses frequency-gradient clustering for sample filtering, while FedHDS already combines local clustering with server clustering of uploaded centroids.

2.The threat model assumes that clean samples are the global majority, but this does not guarantee that clean centroids dominate server clustering. Many small, poison-dominated clients could outnumber larger clean clients, causing the server to select a poisoned reference despite an overwhelming global majority of clean samples.

3.Section 4.3 retains only the cluster containing the global centroid, although HDBSCAN may identify several legitimate clean clusters. This risks removing entire groups of clean instructions, and the aggregate clean accuracy does not establish that minority tasks or data groups are preserved.

Questions for authors

1.What happens when HDBSCAN labels the inserted global centroid as noise? Equation (12) assumes the centroid belongs to a cluster, but it specifies no fallback procedure.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.