acceptodds
Under review as a conference paper at ICLR 2027

VTR-FL:Verifiable Trust Recovery Against Selection-Censored Feedback in Federated Learning

Abstract

Reputation-based client selection in federated learning creates an endogenous feedback problem: once a client is excluded, the system may stop observing the very evidence required to determine whether that client has recovered. We formalize this phenomenon as Selection-Censored Trust (**SCT**) and show that, when post-exclusion observation probability collapses to zero, recovered and persistently malicious trajectories become unidentifiable from system-visible history. We introduce Persistent Trust Observability (**PTO**) and propose a framework named **VTR-FL**, which decouples evidence acquisition from model influence. Low-confidence clients remain observable through a verifiable zero-influence channel, while recent certified evidence, not stale historical reputation, determines whether influence can be progressively restored. A post-commit probation mechanism further constrains same-round role-conditioned exploitation. Under the stated verification and cryptographic assumptions, committed observation propensities and state transitions make recovery decisions challengeable. Across CIFAR-10, Fashion-MNIST, and MNIST, VTR-FL substantially increases recovered-client re-entry and model influence relative to reputation-based selection baselines while preserving comparable task accuracy. Adaptive evaluations further characterize the security boundary of controlled re-entry.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Federated Learning, Client Reputation, Selection-Censored Trust, Trusted Recovery

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

This paper proposes VTR-FL, a federated learning framework addressing selection-censored trust, where excluded clients cannot provide new evidence to recover their reputation. The method separates trust verification from model aggregation through a zero-influence observation channel, recent-evidence-based recovery decisions, and cryptographic commitments for controlled client re-entry. The authors provide theoretical results concerning observability and verification guarantees, alongside experiments on CIFAR-10, Fashion-MNIST, and MNIST reporting improved client recovery and comparable classification accuracy under simulated poisoning scenarios.

Critical strengths
  1. The paper identifies a relevant problem in reputation-based federated learning, where excluding clients can prevent the collection of evidence needed to restore trust after temporary failures.
  2. Separating verification from aggregation is a reasonable design choice, and the experiments demonstrate improved client re-entry across multiple datasets under the evaluated recovery scenarios.
Critical weaknesses
  1. The theoretical novelty is limited. The main results largely formalize intuitive properties, such as the impossibility of detecting recovery without observations and the zero aggregation influence of clients explicitly assigned zero weight. The cryptographic guarantees are similarly conditional on standard assumptions, without establishing substantially new security properties.
  2. The experimental evaluation is overly simplified and does not convincingly demonstrate robustness in realistic adversarial settings. The main recovery experiment uses temporary label corruption that ends at a predetermined round, while the security evaluations cover relatively narrow attack strategies. Comparisons against simpler periodic auditing and re-entry mechanisms under identical verification budgets are missing, making it difficult to justify the complexity of the proposed framework.
  3. The claimed verifiability is not adequately demonstrated in practice. Cryptographic operations and certificate verification are simulated rather than implemented, updates are inspected in plaintext, and experiments rely on a fixed authenticated auditor pool and centralized validation data. These strong assumptions significantly limit the practical relevance of the security claims.
  4. The scalability and performance trade-offs remain insufficiently validated. Experiments use small client populations and relatively simple benchmarks, while verification incurs substantial computation and communication costs. The method also does not consistently outperform competing approaches in classification accuracy, raising questions about its practical advantages.
Questions for authors
  1. What fundamentally new theoretical guarantees does VTR-FL establish beyond standard observability arguments, inverse-propensity estimation, and cryptographic commitment properties?
  2. Can the authors compare against a simpler periodic verification and trust-recovery baseline using identical auditing budgets and recovery criteria?
  3. How would the proposed framework operate without a shared trusted validation pool, fixed authenticated auditors, and access to plaintext client updates? Can the authors demonstrate an actual implementation of the claimed verification protocol?
  4. How does the method perform against persistent adaptive attackers, colluding clients, and larger heterogeneous federations? Can the authors provide more realistic evaluations of security, communication overhead, and accuracy trade-offs?
Reviewer_LrUGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies selection-censored trust in reputation-based federated learning, where excluding a client also removes the observations needed to recognize its subsequent recovery. The authors propose VTR-FL, which separates evidence acquisition from model influence through Verify, Probation, and Train states. The framework combines budget-aware verification scheduling, auditor-certified evidence, propensity-corrected historical statistics, a recent level-and-trend recovery gate, and post-commitment role randomization. It also describes a mechanism for challenging invalid state transitions. Experiments on CIFAR-10, Fashion-MNIST, and MNIST report improved client re-entry and recovered-client influence, supported by recovery-gate ablations, adaptive-attack evaluations, and simulated protocol checks.

Critical strengths
  1. The proposed method allows excluded clients to demonstrate recovery without automatically including their updates in model aggregation.
  2. Figure 2(a) compares three recovery criteria on the same calibration trajectories. Re-entry selectivity increases from 0.017 with historical trust to 0.092 with recent evidence, and further to 0.317 when the recent trend is included. This helps explain why the authors use a separate statistic for recovery.
Critical weaknesses
  1. My main concern is whether the reported metrics measure recovery after exclusion. Appendix D.1 compares the same clients during degradation over rounds [40, 60) and recovery over [60, 100). The recovery window is twice as long, giving clients more opportunities to satisfy the criterion of participating at least once. The global model also changes between these periods, which may affect the validation evidence used to assess clients.
  2. The definition of P(Re)P(\mathrm{Re}) does not clearly state whether a client must have been excluded before it can be counted as having re-entered. The authors should report how many clients were actually excluded and calculate recovery rates for those clients.
  3. To interpret the zero attack exposure in Figure 4, I need a clearer description of what the attacker does when the role is unknown and after promotion.
  4. Table 1 includes an additional adaptive attack, but its poisoning procedure, attack strength, and definitions of ASR and FPR are not sufficiently specified. The method itself also leaves several relevant details unclear.
Questions for authors
  1. How many target clients were actually excluded before the recovery period under each method? Is P(Re)P(\mathrm{Re}) calculated only over these clients? Please also clarify how AUCRD\mathrm{AUC}_{RD} and ΔR\Delta_R are calculated and whether the confirmation study includes clients that remain degraded throughout the same evaluation window.
  2. Please provide the definition of the base aggregation weights, the probation multiplier, and the probabilities used to assign probation roles. If an audit identifies a harmful update, is that update rejected before aggregation, or does the result only affect the client's state in the next round? What bound on probation influence holds after normalization?
  3. Please specify the attacks used in Figure 4 and Table 1, including the attacker’s knowledge, the poisoning operation, the timing of each phase, and the definitions of ASR and FPR.
  4. Does either experiment test an attacker that commits malicious updates before knowing the probation role and resumes attacks after reaching Train? What explains the zero exposure in Figure 4?
Reviewer_q6NnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies the following problem in federated learning: once a malicious client was removed, the system can not know when it becomes trustworthy again. They propose a solution in which clients are still observed but not allowed to influence the model. Experimental results on several datasets show that eliminated clients can be brought back faster than using some reputation baseline.

Critical strengths

None.

Critical weaknesses

The problem of federated learning is very well studied and the paper makes very incremental contributions without practical implications.

The problem is artificial and some of the results obvious. Specifically, why would you want to bring back malicious clients? The paper claims that a system can not know when excluded malicious clients become trustworthy. Well, they were excluded for a reason, so the application does not want to bring them back. This is a policy decision, that the paper treats it as a technical problem. If you do not have any observability channels with the excluded nodes, of course you can not know their behavior. This seems obvious.

The proposed solution seems overly complicated, and consists of a collection of existing mechanisms.

The proposed solution does not solve two fundamental problems: how do auditors distinguish between benign and malicious behavior, the paper seems to rely on this. Second, an attacker can still behave nice and attack once readmitted.

There is no attacker model in the paper.

Finally, the paper uses a lot of very complicated terms that are not standard to refer to simple obvious terms, obfuscating the understanding of the paper.

Questions for authors

How do auditors distinguish between benign and malicious behavior?

Can at attacker still attack immediately when added to the system?

What is your attacker model?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.