acceptodds
Under review as a conference paper at ICLR 2027

Byzantine-Robust Federated Learning with Privacy-Compatible Adaptive Client Selection via Thompson Sampling

Abstract

Byzantine-robust federated learning must decide, in every round, whose updates to admit into aggregation, and this decision has two error modes that pull against each other: rejecting a benign client, and accepting a malicious one. Under non-IID data, benign clients with atypical local distributions are statistically hard to separate from anomalies, so rules that suppress outliers incur persistent false rejections, whereas relaxing them admits attacks optimized to lie inside the benign distribution. Privacy protection tightens the conflict, because the per-update statistics these rules operate on are perturbed, encrypted, or secret-shared; existing defenses respond by rebuilding the detection rule around one specific privacy primitive. We propose BRFL-TS, which changes where the detection evidence comes from: rather than inspecting individual updates, it infers client reliability from the utility of aggregates. Clients are treated as arms of a multiple-play bandit, Thompson Sampling draws client subsets as combinatorial arms, each subset is aggregated into a candidate model that clients score on their own data, and an adaptive reward-and-penalty rule converts subset-level scores into client-level Beta posteriors. Posterior uncertainty keeps early misjudgements refutable, so false rejections decay instead of persisting, while evidence accumulated across rounds exposes stealthy attackers that survive any single round. Because the server never inspects an individual update, the same procedure runs unchanged under differential privacy, homomorphic encryption, and secure multi-party computation. We prove that the active selection domain coincides with the benign client set—no benign client rejected and no Byzantine client accepted—in all but an fraction of the training rounds, and show empirically that BRFL-TS keeps both error rates low across eight Byzantine attacks, increasing data heterogeneity, and all three privacy primitives.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Robustness, Trustworthiness, Privacy

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_wggLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses client reliability misjudgments caused by data heterogeneity and participation bias arising from historical trust assessments in Byzantine-robust federated learning. It proposes BRFL-TS, an adaptive client selection method based on Thompson Sampling. The method maintains Beta trust distributions for clients to construct candidate client groups, evaluates group utility using the local evaluation results of candidate aggregated models, and assigns rewards and penalties based on client membership across groups with different utility levels. Client trust and subsequent selection probabilities are dynamically updated, while model reevaluation and rollback mechanisms are introduced to limit the impact of low-utility updates. Theoretically, the paper establishes a lower bound on the probability of receiving renewed trust feedback within finite training windows and an upper bound on the expected waiting time under continued training. Experiments across four datasets and eight untargeted poisoning attacks evaluate the robustness of BRFL-TS under varying data heterogeneity, Byzantine client ratios, and DP, HE, and MPC update protection settings. Client-level decision metrics and trust evolution analyses further demonstrate the effectiveness of the proposed method.

Critical strengths
  1. Utility-based reliability assessment reduces dependence on individual model-update statistics.
  2. The adaptive credit assignment mechanism supports reliability reassessment without permanently excluding penalized clients.
  3. The experimental evaluation provides evidence of robustness and correction of early reliability misjudgments.
Critical weaknesses
  1. BRFL-TS incorporates multiple components, including Thompson Sampling, adaptive rewards and penalties, cross-group membership ranking, and model rollback. However, it lacks systematic ablation studies, making it difficult to determine each component's independent contribution to overall performance. In particular, direct comparisons between the adaptive reward-and-penalty mechanism and simpler strategies are missing.
  2. Group-level model utility is used to assess client reliability, but group performance does not directly reflect individual contributions. Under Non-IID settings, differences in data distributions and interactions among clients further complicate credit assignment. The existing experiments lack direct validation of the attribution accuracy of this mechanism.
  3. BRFL-TS requires constructing multiple candidate models in each training round and having all clients evaluate them. The reported upper bound on plaintext-equivalent communication overhead is approximately 6.5 times that of a full-participation baseline. However, empirical measurements of actual communication volume, computational costs, and runtime are not provided.
  4. The existing experiments primarily report model accuracy under DP, HE, and MPC settings, but a comprehensive evaluation of the end-to-end privacy-preserving protocol is lacking. The paper has not sufficiently analyzed privacy guarantees, runtime efficiency, and system scalability in practical deployment.
Questions for authors
  1. Could the authors provide ablation studies of the key components to clarify their individual contributions? What are the respective effects of these components on model accuracy and client-level misclassification rates?
  2. Does the client-level credit assignment mechanism remain effective under more severe data heterogeneity? There is no direct correspondence between group-level model utility and individual client contributions, and Non-IID data distributions and interactions among clients may affect credit assignment. Could the authors further evaluate the accuracy of reward-and-penalty decisions under different levels of data heterogeneity and examine the relationship between group utility and individual contributions?
  3. Limited theoretical support for correcting early client misclassification. Theorem 1 establishes a positive lower bound on the probability that a client receives renewed trust feedback within a finite training window. However, this result does not directly guarantee that an initially misclassified benign client will recover its trust. Could the authors further clarify how this theoretical result supports the claim of mitigating the Matthew effect in client participation?
  4. Could the authors further clarify the end-to-end privacy guarantees under DP, HE, and MPC settings? The existing results primarily demonstrate model accuracy under different privacy protection mechanisms, but the privacy guarantees of the complete protocol have not been sufficiently analyzed. Could the distribution of candidate models and the reporting of client evaluation scores expose additional information? How is such information protected under different privacy-preserving settings?
7 comments
AuthorsOpenReview ↗

Response to Question 1: Additional information regarding ablation studies

Thank you for your comments on our paper. We conducted an ablation study of BRFL-TS under the PoisonedFL attack using CIFAR-10 and AlexNet. All configurations used a non-IID Dirichlet data partition (β=0.5\beta=0.5), with 20 clients, 20% of which were Byzantine, and ran for 100 communication rounds. In each round, we formed 10 candidate groups of 3 clients each. We used the same three random seeds (42, 123, and 3407) for every configuration.

For each run, we report model test accuracy and two client-level error rates. The false rejection rate (FRR) is the proportion of explicit decisions on benign clients that assign a penalty. The false acceptance rate (FAR) is the proportion of explicit decisions on Byzantine clients that assign a reward.

Table 1. Ablation study of BRFL-TS on CIFAR-10 under PoisonedFL. Values are percentages reported as mean ±\pm sample standard deviation across three random seeds.

\text{Configuration} & \text{Model test accuracy} & \text{FRR} & \text{FAR} \cr \text{Full BRFL-TS} & 79.97\pm1.26 & 6.84\pm1.62 & 8.59\pm3.55 \cr \text{Uniform sampling} & 79.46\pm1.32 & 11.32\pm0.98 & 7.69\pm2.42 \cr \text{Fixed penalty quota} & 79.84\pm1.40 & 33.27\pm1.68 & 5.75\pm2.02 \cr \text{Random assignment} & 79.63\pm0.90 & 8.76\pm1.52 & 8.31\pm1.42 \cr \text{No rollback} & 79.94\pm0.82 & 8.44\pm1.87 & 5.55\pm0.90 \end{array}$$ #### Thompson sampling Replacing trust-weighted group sampling with uniform sampling increased FRR by 4.48 percentage points and decreased model test accuracy by 0.51 percentage points. The most notable benefit of Thompson sampling was a lower rate of erroneous penalties on benign clients, reflecting its ability to correct early misclassifications. #### Adaptive penalty quota Fixing the penalty quota at two per round increased FRR by 26.43 percentage points and decreased model test accuracy by 0.13 percentage points. In the full method, the adaptive rule assigned an average of 0.46 penalties per round. This comparison shows that the adaptive quota had the largest effect on the false rejection of benign clients. As client selection progressively narrows and filters out malicious clients, fewer malicious clients remain in the lowest-ranked group. These results demonstrate that the adaptive reward-and-penalty mechanism adjusts the penalty quota in response to this change, thereby reducing false rejection. #### Lexicographic penalty assignment Replacing lexicographic assignment with random selection of penalty recipients from the lowest-ranked group increased FRR by 1.92 percentage points and decreased model test accuracy by 0.34 percentage points. The average penalty quotas were similar (0.44 with random assignment and 0.46 in the full method). This suggests that using membership information from utility-ranked groups helps identify malicious clients to penalize more accurately at a comparable penalty budget, thereby limiting the false rejection of benign clients. #### Rollback Disabling rollback increased FRR by 1.60 percentage points and decreased model test accuracy by about 0.03 percentage points, yielding accuracy close to that of the full method. In the full method, rollback was triggered an average of 0.33 times per run. It also occurred in other ablation configurations. Rollback primarily serves as a safeguard against cases where a malicious update in the highest-ranked group substantially degrades model accuracy. Its effect is less pronounced when the other mechanisms are functioning as intended. Under this configuration, the complete BRFL-TS scheme achieves the lowest average FRR and the highest average model test accuracy. As for why the average FAR remains largely similar across the various ablation variants, this is because the optimal group is selected based on model utility, and robust aggregation—from which the FAR is calculated—is performed using client updates from this optimal group. Since removing any of the aforementioned modules does not fundamentally alter this process, the FAR remains within essentially the same range.
AuthorsOpenReview ↗

Response to Question 2: Supplementary Information Regarding the Experiment on the Effectiveness of the Credit Allocation Mechanism

We evaluated client-level credit assignment in BRFL-TS under the PoisonedFL attack to assess its effectiveness under stronger data heterogeneity. We used AlexNet on CIFAR-10 with two non-IID Dirichlet partitions (β=0.5\beta=0.5 and 1.01.0). Each run involved 16 benign and 4 Byzantine clients over 100 training rounds. We focused on benign clients whose first penalty occurred in rounds 1–10.

Decision coverage

To assess how well the credit assignment mechanism revisits early penalties on benign clients, we measured the proportion of these clients that received at least one reward within 10, 25, and 50 rounds of their first penalty. As shown in Table 1, all such clients at β=1.0\beta=1.0 received a reward within 10 rounds. Within 25 rounds, every client in the cohort had received a reward, including the client at β=0.5\beta=0.5. These results show that the mechanism can promptly revise its trust assessments in response to new model utility evidence. To further assess this revision, Table 2 reports the clients' final B/FB/F ratios. The posterior mean trust increased markedly for all four clients. These clients also accumulated penalties after round 10. Under heterogeneous data, their updates may perform poorly in some rounds, leaving them eligible for another penalty if they appear in the lowest-ranked group. This behavior is consistent with the design of the decision rule.

Table 1. Coverage of subsequent rewards among benign clients penalized early. Each cumulative proportion uses the full early-penalty cohort in its row as the denominator.

\beta & \leq 10 & \leq 25 & \leq 50 & \text{By round 100} \cr 0.5 & 0/1 & 1/1 & 1/1 & 1/1 \cr 1.0 & 3/3 & 3/3 & 3/3 & 3/3 \cr \text{Pooled} & 3/4 & 4/4 & 4/4 & 4/4 \end{array}$$ #### Trust and recovery-time distribution We measured the waiting time from each client's first erroneous penalty to its first subsequent reward, defined as the first subsequent round in which its reward counter $B$ was incremented. A client's contribution was considered retained when its update was included in an aggregate that passed the rollback check and became the global model used for subsequent training. Both runs had zero rollbacks, so each client's first subsequent reward coincided with its first retained contribution after the penalty. Table 1 reports the cumulative results, and Table 2 gives the recovery timeline for each client. The median waiting time was 7 rounds, with a range of 4–17 rounds. The last recovery occurred at round 26. Table 2 compares each early-penalized client's cumulative reward-to-penalty ratio $B/F$ at the end of rounds 10 and 100. Under the $\mathrm{Beta}(B,F)$ distribution, a higher $B/F$ corresponds to a higher posterior mean trust $B/(B+F)$. This ratio increased markedly for all four clients. At round 100, the sole early-penalized benign client at $\beta=0.5$ had a lower $B/F$ ratio than the three clients at $\beta=1.0$. This observation is consistent with the possibility that stronger heterogeneity can make some benign updates less useful in certain group compositions, making those clients more likely to appear in the lowest-ranked group and receive penalties that lower their posterior mean trust. These results suggest that the credit assignment mechanism can use group utility to revise client-level trust even under heterogeneous data. **Table 2.** Individual recovery times and paired $B/F$ ratios for the early-penalty cohort. Recovery times are in rounds. Each fraction shows the original reward and penalty counters. $$\small\begin{array}{crrrrrr} \beta & \begin{array}{c}\text{Client}\cr \text{ID}\end{array} & \begin{array}{c}\text{First}\cr \text{penalty}\end{array} & \begin{array}{c}\text{Next}\cr \text{reward}\end{array} & \text{Wait} & \begin{array}{c}B/F\cr \text{Round 10}\end{array} & \begin{array}{c}B/F\cr \text{Round 100}\end{array} \cr 0.5 & 12 & 9 & 26 & 17 & 1/2 & 8/6 \cr 1.0 & 4 & 7 & 11 & 4 & 1/2 & 9/6 \cr 1.0 & 18 & 1 & 11 & 10 & 1/2 & 14/4 \cr 1.0 & 19 & 2 & 6 & 4 & 4/2 & 14/6 \end{array}$$ #### False rejection rate (FRR), false acceptance rate (FAR), and final test accuracy Table 3 reports the FRR, FAR, and final CIFAR-10 test accuracy under PoisonedFL. Both decision error rates varied only slightly across the two non-IID settings. Final test accuracy was 81.94% at $\beta=0.5$ and 83.36% at $\beta=1.0$, a difference of 1.42 percentage points. The similar decision error rates and final test accuracies support the robustness of BRFL-TS across the two tested levels of heterogeneity. **Table 3.** False rejection rate, false acceptance rate, and final CIFAR-10 test accuracy at round 100 under PoisonedFL. All values are percentages. $$\begin{array}{lrrr} \beta & \text{FRR} & \text{FAR} & \text{Final test accuracy} \cr 0.5 & 7.23 & 7.94 & 81.94 \cr 1.0 & 7.55 & 9.09 & 83.36 \end{array}$$
AuthorsOpenReview ↗

Response to Question 3: A Supplement Regarding the Correction of Early Client-Side Misjudgment Theories

Thank you for pointing out the gap between “receiving feedback again” and “recovering from an erroneous early penalty.” Our motivation for this proof is that selection rules that favor clients with “strong historical performance” may repeatedly select the same clients, leaving insufficiently explored clients with few opportunities to build a performance history. We therefore wanted to show that, in BRFL-TS, Thompson sampling still gives a client a chance to enter a candidate group even after poor early feedback. The theoretical bound on the probability of feedback offers limited support for this conclusion. We discussed this issue in depth and believe that the underlying reason is that, among the multiple candidate models in our method, clients in groups with intermediate scores receive no feedback. The proof therefore provides a lower bound on the probability of feedback only for the case in which the client appears in every group. In practice, whether a client receives feedback is closely tied to its actual performance, a relationship that the theoretical analysis did not account for. We therefore drew on the ideas of Good Arm Identification [1] and composite anonymous feedback [2] to formulate a utility separation assumption and derive a new lower bound on the probability of feedback.

First, we derive a lower bound on the probability that a client is selected in QQ draws without replacement in a single round. Given client ii's Thompson sample xx, if the first rr draws all miss ii, the total remaining weight of the other clients is at most K−1−rK-1-r. Multiplying the upper bounds on the probability of missing ii at each draw and taking the complement gives the following lower bound on the probability that a group of size QQ contains ii.

ℓ(x)=1−∏r=0Q−1K−1−rK−1−r+x.\ell(x)=1-\prod_{r=0}^{Q-1}\frac{K-1-r}{K-1-r+x}.

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

This paper proposes BRFL-TS, a Byzantine-robust federated learning framework that assesses client reliability through sampled group-model utility rather than individual plaintext updates. Beta-based sampling guides group selection, while median client evaluations and adaptive credit assignment update trust. A rollback mechanism controls model retention. The theoretical analysis bounds the probability and waiting time for renewed feedback and conditionally relates group purity to benign-client coverage. Experiments across four datasets, eight attacks, and several update-protection settings report improved accuracy and favorable client-level decisions, including an 8.01 percentage-point gain on Fashion-MNIST. These benefits require additional evaluation and communication, with an estimated communication upper bound of 6.5$\times$ a full-participation baseline for M=10M=10, excluding cryptographic overhead.

Critical strengths
  1. The paper addresses the difficulty of distinguishing heterogeneous benign updates from malicious ones through individual-update screening. Evaluating candidate aggregates offers a coherent alternative, while Thompson-weighted sampling preserves opportunities for reassessment after negative feedback. Adaptive rewards and penalties, reward-set reevaluation, and rollback are integrated into a complete procedure across training rounds.

  2. The evaluation covers multiple datasets and attack families, varies data heterogeneity and Byzantine ratios, and examines both model accuracy and client-level decision errors. Experiments with falsified scores and selected trust trajectories provide additional insight into the method’s behavior. The reported results are promising, particularly on Fashion-MNIST. The explicit plaintext-equivalent communication accounting and acknowledgment of unmeasured computational and cryptographic costs help clarify the practical trade-offs.

Critical weaknesses
  1. Privacy compatibility requires a concrete protocol and leakage analysis. Equations (3), (8), and Algorithm 1 release overlapping subset aggregates without sufficiently specifying decryption permissions, collusion assumptions, or subset-specific masking. Singleton reward sets are also permitted; releasing a recoverable singleton model under noiseless HE/MPC reveals that client’s update to parties knowing the current model. For DP, the neighboring relation, privacy accounting, resulting (ϵ,δ)(\epsilon,\delta), and protection of evaluation scores remain unspecified.

  2. The theory establishes renewed feedback, not correction of erroneous judgments. Theorem 1 counts both rewards and penalties, so repeated punishment satisfies its guarantee without ensuring trust recovery or retained contributions. Its expected-waiting-time upper bound evaluates to approximately 1.13×10141.13\times10^{14} rounds at initialization for K=20,M=10K=20,M=10, making it uninformative for explaining reassessment within the 100-round experiments. Theorem 2 remains conditional without an informative algorithm-specific regret bound. Recovery-time and post-rollback contribution statistics across independent runs would better substantiate practical reassessment.

  3. The robustness evaluation and comparisons are insufficiently controlled. Malicious-scoring experiments cover only score reversal and random reports under IID Sign Flipping, although heterogeneous honest scores can allow coordinated manipulation of median rankings. Targeted non-IID attacks are needed. Ablations should also isolate sampling, adaptive penalties, lexicographic assignment, and rollback. Across-attack standard deviations do not establish seed-level reliability, while differing decision definitions, participation budgets, and access to evaluation feedback limit baseline comparability.

  4. Reproducibility and the Tiny-ImageNet configuration require clarification. Important training, attack, validation, and replication details are missing. Standard Tiny-ImageNet has 200 classes, whereas Table 9 repeatedly reports approximately 10% accuracy alongside a 94.20% attack-free reference. These observations do not prove an error, but warrant explicit specification of class count, subset construction, label mapping, evaluation metric, data splits, and pretrained initialization before the results can be interpreted reliably.

Questions for authors
  1. Which concrete DP, CKKS, and secure-aggregation/MPC protocols instantiate P\mathcal P, and what are their collusion, decryption, and output-visibility assumptions? How is individual-update disclosure prevented for singleton reward sets and overlapping aggregates? For DP, please clarify whether evaluation scores are protected, specify the neighboring relation and privacy accounting, and report the resulting (ϵ,δ)(\epsilon,\delta).

  2. Can the authors provide per-attack, multi-seed results with uncertainty estimates and complete experimental configurations? Please clarify comparison budgets, including participation, evaluation access, communication, and computation. Please also specify Tiny-ImageNet’s classes, samples, splits, label mapping, and pretrained initialization, and explain the repeated approximately 10% results in Table 9.

  3. What fraction of mistakenly penalized benign clients subsequently receive rewards and contribute to the retained model within the training horizon? Please report recovery-time distributions and decision coverage alongside FRR/FAR. Can component ablations and targeted, coordinated score/update attacks under non-IID data establish which mechanisms drive the gains and when aggregate-utility feedback remains reliable?

9 comments
AuthorsOpenReview ↗

Response to Question 1:Supplementary Clarification Regarding the Privacy Policy

Thank you for pointing out the issues with the implementation details of our privacy schemes. Our study focused primarily on whether these schemes could be deployed with our method and left the level of privacy they provide in practice unexamined. We describe the specific protocols below.

In the HE scheme, we assume that a trusted party holds the secret key and that clients use the public key for encryption. The attacker aims to obtain model updates and carry out a model inversion attack to recover client information. In each round, each selected client computes its local update Δi\Delta_i, encrypts the full update as Encpk(Δi)\mathrm{Enc}_{pk}(\Delta_i), and sends the ciphertext to the server. The server uses homomorphic weighted sums to construct all candidate aggregates and the aggregate for the reward set, and broadcasts the encrypted candidate models to the clients. The trusted party decrypts the candidate models and sends them to all clients for evaluation. After completing their evaluations, the clients send their results to the server. The server tallies the evaluation results, assigns rewards and penalties to the clients, and performs aggregation using a homomorphic weighted sum. Throughout this process, HE protects individual client model updates, making it difficult for the attacker to obtain model information unless they obtain the secret key and use it to decrypt the updates.

The MPC scheme protects privacy using Bonawitz-style additive masking. For the privacy protocol, we assume a semi-honest attacker who follows the protocol but records intermediate messages in an attempt to infer private client information. At the start of training, the server selects clients, forms multiple client groups, and broadcasts the group assignments. Clients within each group coordinate their masks, apply the masks to their local model updates, and send the masked updates to the server. The server aggregates the client model updates within each group to form a candidate model and sends the candidate models to the clients for evaluation. The clients send their evaluation scores to the server, which likewise assigns rewards and penalties and performs robust aggregation based on the evaluation results. The server and clients obtain multiple plaintext candidate models. The sets of client updates aggregated into these models may overlap, making it possible to construct and solve a system of equations for the client updates. This risk can be avoided by limiting MM or QQ, or by having the server check the sampled groups for privacy risks in advance.

For differential privacy, we use client-level replace-one adjacency with a fixed set of registered clients. Two inputs are adjacent if the full local training dataset of one registered client is replaced. We again assume a semi-honest attacker who attempts to infer private information from model updates. The scheme clips each full update to satisfy ∥Δi∥2≤Cu\lVert\Delta_i\rVert_2\le C_u, giving a sensitivity of at most 2Cu2C_u. Each client draws Gaussian noise once and uploads the resulting protected local update. All candidate groups in the same round reuse that protected update. Applying differential privacy to evaluation scores falls outside the scope of our threat model.

When the reward set contains only one client or multiple aggregation subsets overlap, updates protected by differential privacy or HE meet the basic privacy requirements under our threat model. We addressed overlapping aggregation subsets for MPC in the protocol description above. The reward set can indeed contain a single client, in which case exposing that client's update directly to the server poses a privacy risk. The reward set is formed by removing the penalized clients in the lowest-scoring group from the highest-scoring group. These two groups typically have little overlap, since substantial overlap would make it difficult for them to differ in utility. A singleton reward set is therefore a special case and also indicates that the highest-scoring group did not have a substantial advantage in utility. In this case, we could skip the current round and proceed to the next training round.

AuthorsOpenReview ↗

Response to Question 2: Supplementary Experiments Regarding Multiple Attacks

We provide the per-attack results and experimental configuration below.

Per-attack results

We evaluated BRFL-TS on CIFAR-10 under eight attacks using random seeds 42, 123, and 3407. All 24 runs completed. For each run, we computed the global model's mean test accuracy over rounds 91–100 and recorded its test accuracy at round 100. Table 1 reports the last-10-round mean for each seed and the mean and sample standard deviation (SD, with denominator n−1n-1 and n=3n=3) across seeds for each attack.

Table 1. Per-attack CIFAR-10 test accuracy (%). The three seed columns report the mean accuracy over rounds 91–100. Both summary columns give mean ±\pm sample SD across the three seeds.

\text{Attack} & \begin{array}{c}\text{Seed}\cr 42\end{array} & \begin{array}{c}\text{Seed}\cr 123\end{array} & \begin{array}{c}\text{Seed}\cr 3407\end{array} & \begin{array}{c}\text{Last 10}\cr \text{rounds}\end{array} & \begin{array}{c}\text{Round}\cr 100\end{array} \cr \text{Sign Flipping} & 80.72 & 76.12 & 82.18 & 79.67\pm3.16 & 79.74\pm3.25 \cr \text{Label Flipping} & 80.20 & 78.72 & 81.91 & 80.28\pm1.60 & 81.29\pm1.96 \cr \text{Fang-v1} & 78.33 & 77.92 & 80.27 & 78.84\pm1.26 & 79.02\pm2.04 \cr \text{Fang-v2} & 81.60 & 80.23 & 81.31 & 81.05\pm0.72 & 81.04\pm1.95 \cr \text{I-FMPA} & 80.52 & 80.02 & 81.71 & 80.75\pm0.87 & 81.96\pm2.39 \cr \text{Min-Max} & 78.73 & 78.61 & 80.74 & 79.36\pm1.19 & 81.50\pm1.67 \cr \text{Min-Sum} & 77.81 & 78.55 & 80.06 & 78.81\pm1.15 & 79.13\pm1.17 \cr \text{PoisonedFL} & 80.87 & 79.01 & 79.90 & 79.93\pm0.93 & 80.55\pm0.92 \end{array}$$ Across the eight attacks, mean test accuracy over the final 10 rounds ranged from 78.81% to 81.05%, demonstrating the robustness of BRFL-TS across the attacks tested. #### Experimental configuration We used AlexNet with a non-IID Dirichlet data partition with concentration parameter $\beta=0.5$. There were $K=20$ clients (16 benign and 4 Byzantine) and $T=100$ global training rounds. In each round, BRFL-TS formed $M=10$ candidate groups with $Q=3$ distinct clients per group. The rollback threshold was $0.05$. For local training on CIFAR-10, we used SGD for three local epochs with momentum $0.9$ and weight decay $0.001$. #### Participation and evaluation access Under the BRFL-TS protocol, let $U_t=\left|\bigcup_{m=1}^{M} S_{m,t}\right|$ denote the number of distinct clients selected for training in round $t$. Each selected client trains once and uploads one update, which is reused across all candidate groups to which it belongs. All clients, including those that did not train in the round, receive every candidate model and score it on their local data. The server ranks the candidates by median client score. If the reward set coincides with the winning group, the server reuses the winning model's score. If the reward set is nonempty and differs from the winning group, the server requests one additional score from each client. #### Communication and computation Let $I_t\in\lbrace0,1\rbrace$ indicate whether the model aggregated from the reward set is reevaluated. Counting plaintext point-to-point payloads, with $B$ bytes per model or update and $s$ bytes per scalar score, the per-round communication cost of BRFL-TS is $$C_t=\bigl[2U_t + 20(10+I_t)\bigr]B + 20(10+I_t)s \le 260B+220s.$$ Under the same accounting, if $n_t$ clients each download one model and upload one update per round, the communication cost is $2n_tB$. With negligible score payloads, the BRFL-TS communication upper bound is approximately $6.5\times$ the full-participation baseline ($40B$). Each round involves $U_t$ local training jobs, $200+20I_t$ client-side model evaluations, and server-side construction of 10 candidate aggregates plus at most one reward-set aggregate.
AuthorsOpenReview ↗

Response to Question 2: Supplementary Clarification Regarding Experimental Results on the Tiny-ImageNet Dataset

Below are additional details about the Tiny-ImageNet experiments. Given our limited experimental resources, training ViT-B/16 on the full 200-class dataset would have been too demanding for us. Therefore, in all of our Tiny-ImageNet experiments, we used a fixed 10-class subset of Tiny-ImageNet-200. We will revise the relevant parts of the manuscript to make this explicit.

Classes, samples, and label mapping

We constructed the subset from the officially released Tiny-ImageNet-200 dataset (tiny-imagenet-200.zip). Each class is identified by its WordNet ID (wnid). We selected the first 10 classes in lexicographic order by wnid.

\text{New label} & \mathtt{wnid} & \text{Synset (from }\text{words.txt}\text{)} \cr 0 & \mathtt{n01443537} & \text{goldfish} \cr 1 & \mathtt{n01629819} & \text{European fire salamander} \cr 2 & \mathtt{n01641577} & \text{bullfrog} \cr 3 & \mathtt{n01644900} & \text{tailed frog} \cr 4 & \mathtt{n01698640} & \text{American alligator} \cr 5 & \mathtt{n01742172} & \text{boa constrictor} \cr 6 & \mathtt{n01768244} & \text{trilobite} \cr 7 & \mathtt{n01770393} & \text{scorpion} \cr 8 & \mathtt{n01774384} & \text{black widow} \cr 9 & \mathtt{n01774750} & \text{tarantula} \end{array}$$ For each selected class, we retained **all official images** for that class. - **Training set** — 500 images per class $\to$ **5,000** training images in total - **Validation set / global test set** — 50 images per class $\to$ **500** evaluation images in total The original Tiny-ImageNet labels were mapped one-to-one (injectively) to $\lbrace0,\ldots,9\rbrace$ in the order given above. The validation images were filtered to retain only these 10 classes, and the same mapping was used to convert their labels. Therefore, the classification head has dimension **10**, and model accuracy is computed on this 10-class classification task. #### Data partitioning across clients The Tiny-ImageNet-10 configuration is as follows. $$\small\begin{array}{ll} \text{Item} & \text{Setting} \cr \text{Number of clients (}K\text{)} & 20 \cr \text{Global rounds (}T\text{)} & 30 \cr \text{Candidate groups (}M\text{)} & 10 \cr \text{Clients per group (}Q\text{)} & 4 \cr \text{Byzantine proportion} & \text{0.2 (unless otherwise specified)} \cr \text{Batch size} & 32 \cr \text{Local epochs} & 1 \cr \text{Optimizer / learning rate} & \begin{array}{l}\text{AdamW, }1\times10^{-4}\cr \text{Weight decay 0.01}\end{array} \cr \text{Client data partition} & \begin{array}{l}\text{Independent and identically}\cr \text{distributed (IID), or non-IID}\cr \text{under a Dirichlet distribution}\cr \text{(see the respective tables, }\beta\in\lbrace1.0,0.5\rbrace\text{)}\end{array} \end{array}$$ #### Preprocessing and pretrained initialization The images were resized to **$224\times224$** and preprocessed using the Hugging Face image processor for `google/vit-base-patch16-224` (using ImageNet-based means and standard deviations). We used **ViT-B/16** as the backbone and initialized it by loading the same pretrained checkpoint (`google/vit-base-patch16-224`, or its local mirror directory `models/vit-base-patch16-224`). Because the official checkpoint contains a 1000-class output head, we replaced it with a **new 10-class classification head** by setting `num_labels=10` and `ignore_mismatched_sizes=True`. This new classification head was randomly initialized. The pretrained Transformer weights were retained and fine-tuned during federated learning (FL). All layers in the model remained unfrozen. #### Why the values in Table 9 are repeatedly around 10%, while the attack-free reference is around 94% First, the repeated values of around 10% in Table 9 appear only at the strong differential privacy levels G3 and G4. At these levels, the noise substantially disrupts training, resulting in values such as 10.00% or $(10.05\pm0.28)$% (for example, the Fed-NGA results in Table 9 at the strong differential privacy (DP) levels G3/G4). Under conditions without privacy protection and at the weak differential privacy levels, our method consistently achieves higher model accuracy than the other defense methods, demonstrating its compatibility with differential privacy. In particular, without differential privacy, our method approaches the attack-free baseline (i.e., FedAvg with no defense and no attack), and model accuracy can exceed 90%. This shows that the selected subset, label mapping, and pretrained initialization can achieve excellent performance when the model does not “collapse.” The above 10% accuracy results are simply a normal outcome under strong noise interference.
Reviewer_ffE7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes BRFL-TS, a Byzantine-robust federated learning framework that avoids inspecting individual client updates and instead evaluates the utility of candidate models aggregated from sampled client subsets. It translates relative utility among these candidate aggregates into client-level reward/penalty updates to Beta trust distributions, which in turn guide Thompson-weighted subset sampling in subsequent rounds, while a rollback rule prevents harmful updates from being adopted. The authors analyze “reassessment opportunities” by proving a positive lower bound on the probability of renewed client feedback within a finite training window and provide empirical results across four datasets and eight untargeted attacks (with additional experiments under DP/HE/MPC interfaces), showing improved model utility and low false rejection/acceptance rates.

Critical strengths
  • The shift from per-update inspection to utility-based evaluation of group-aggregated candidate models is a privacy-compatible design choice.

  • The adaptive credit assignment that leverages membership across utility-ranked groups to determine rewards/penalties is an interesting and original mechanism.

  • Thompson-weighted subset sampling integrated with adaptive trust updates is a principled way to balance exploitation with continued exploration.

Critical weaknesses
  • The main theoretical guarantee is likely meaningless in realistic regimes: the per-round feedback lower bound scales as 1/KM1/K^M (Eq. 11), which is small for typical K,MK, M, yielding expected waiting times far beyond practical training times. This weakens the claim that prior penalties do not eliminate reassessment opportunities in practice.

  • The credit-assignment quota κ\kappa relies on Beta samples and ad-hoc Bernoulli trials, and its statistical calibration and robustness to colluding adversaries who manipulate evaluation scores (even if median is used) is under-analyzed.

  • Critical implementation details for HE/MPC compatibility and key management are not specified.

  • The overhead from distributing and evaluating multiple candidate models is very high and scalability of the method remains unclear.

  • There is no evaluation on backdoor attacks.

Questions for authors
  • Practicality of the theoretical bound: For typical KK and MM (e.g., K=100K=100, M=10M=10), ci(t)≈1/KMc_i(t) ≈ 1/K^M is small. Can you provide tighter, practically meaningful bounds (e.g., on inclusion probability or feedback return over HH rounds) that reflect your weighted-sampling-without-replacement and actual parameter ranges?

  • DP/HE/MPC integration: In HE or MPC settings, who holds the decryption keys? How does the server broadcast MM encrypted candidate models for all clients to recover without compromising secure aggregation or violating key-distribution assumptions? Please detail a concrete protocol and its trust/attack surface.

4 comments
AuthorsOpenReview ↗

Response to Question 1: On Further Optimization within Theoretical Circles

Thank you for raising this point. We agree that the theoretical bound yields a feedback probability that is too small to be practically meaningful under the experimental parameter settings. After examining this issue in more detail, we believe the main reason is that, among the multiple candidate models in our method, clients in groups with intermediate scores do not receive feedback. The proof therefore provides a feedback lower bound only for the case in which a client appears in every group. In practice, whether a client receives feedback is closely tied to its actual performance, which the theoretical analysis did not account for. To address this, we drew on the ideas of Good Arm Identification [1] and composite anonymous feedback [2] to formulate a utility separation assumption and derive a new lower bound on the feedback probability.

First, we derive a lower bound on the probability that a client is selected in a sequence of QQ draws without replacement. Given client ii's Thompson sample xx, if the first rr draws miss ii, the total remaining weight of the other clients is at most K−1−rK-1-r. Multiplying the upper bounds on the miss probabilities across the draw positions and taking the complement gives the following lower bound on the probability that a group of size QQ contains ii.

ℓ(x)=1−∏r=0Q−1K−1−rK−1−r+x.\ell(x)=1-\prod_{r=0}^{Q-1}\frac{K-1-r}{K-1-r+x}.

Since the MM candidate groups are drawn independently, averaging over XX gives the lower bound a=E[1−(1−ℓ(X))M]a=\mathbb{E}[1-(1-\ell(X))^M] on the probability that the client is selected in at least one of the MM draws.

To derive a lower bound on the probability of feedback after the client is selected, we now define the utility separation assumption. Given the round history, for a high-utility client we assume that a group containing ii has a lower score bound above the high threshold with probability at least α\alpha, and that a group excluding ii has an upper score bound above that threshold with probability at most β\beta. For a low-utility client, we use a low threshold and require that a group containing ii has an upper score bound below the threshold with probability at least α\alpha, and that a group excluding ii has a lower score bound below the threshold with probability at most β\beta. Client ii receives feedback if and only if the utility of at least one group containing ii and the utilities of all of the at most M−1M-1 groups excluding ii lie on opposite sides of the threshold. We therefore obtain the following lower bound on the probability that client ii receives feedback.

λi=aα(1−β)M−1ρi.\lambda_i=a\alpha(1-\beta)^{M-1}\rho_i.

Here, ρi\rho_i is the probability that client ii receives penalty feedback in the lowest-scoring group. Our adaptive reward and penalty mechanism adjusts the number of penalties assigned within that group, so a client in the lowest-scoring group receives penalty feedback with the corresponding probability. In the highest-scoring group, client ii is guaranteed to receive feedback, so we set ρi=1\rho_i=1. While ii waits for its next participation or feedback, its Beta counts remain fixed. If the score and penalty conditions hold over the relevant waiting histories, the probability of participating at least once within HH rounds is at least 1−(1−a)H1-(1-a)^H, and the probability of receiving feedback at least once is at least 1−(1−λi)H1-(1-\lambda_i)^H.

We consider K=100K=100, M=10M=10, which you have mentioned, and set Q=20Q=20. Suppose that client ii has high utility but is misclassified and penalized for 10 consecutive rounds. Its state at the end of round 10 is then Beta(1,11)\mathrm{Beta}(1,11). We use this extreme case to assess the practical relevance of the feedback probability. In practice, the client would have more opportunities to receive feedback. Under these conditions, the lower bound on the probability that client ii is selected in one training round is approximately 15.91%, and the lower bound on the probability that it is selected within the next 10 rounds is 82.32%. With illustrative score conditions α=40\alpha=40% and β=5\beta=5%, and with ρi=50\rho_i=50% for the low-utility case, the lower bounds over the remaining 90 rounds are 97.49% for receiving reward feedback and 83.85% for receiving penalty feedback.

References

  1. Julian Katz-Samuels and Kevin Jamieson. The True Sample Complexity of Identifying Good Arms. In Proceedings of the 23rd International Conference on Artificial Intelligence and Statistics, PMLR 108, pp. 1781–1791, 2020.
  2. Nicolò Cesa-Bianchi, Tommaso Cesari, Roberto Colomboni, Claudio Gentile, and Yishay Mansour. Nonstochastic Bandits with Composite Anonymous Feedback. Journal of Machine Learning Research, 23(277):1–24, 2022.
AuthorsOpenReview ↗

Response to Question 2:Regarding Specific Protocols for HE/MPC Privacy Solutions

Below, we provide the detailed protocol for the privacy scheme, along with its trust assumptions and attack surface.

In the HE scheme, we assume that a trusted party holds the secret key and that clients use the public key for encryption. The attacker aims to obtain model updates and carry out a model inversion attack to recover client information. In each round, each selected client computes its local update Δi\Delta_i, encrypts the full update as Encpk(Δi)\mathrm{Enc}_{pk}(\Delta_i), and sends the ciphertext to the server. The server uses homomorphic weighted sums to construct all candidate aggregates and the aggregate for the reward set, and broadcasts the encrypted candidate models to the clients. The trusted party decrypts the candidate models and sends them to all clients for evaluation. After completing their evaluations, the clients send their results to the server. The server tallies the evaluation results, assigns rewards and penalties to the clients, and performs aggregation using a homomorphic weighted sum. Throughout this process, HE protects individual client model updates, making it difficult for the attacker to steal model information unless they obtain the secret key and use it to decrypt the updates.

The MPC scheme protects privacy using Bonawitz-style additive masking. For the privacy protocol, we assume a semi-honest attacker who follows the protocol but records intermediate messages in an attempt to infer private client information. At the start of training, the server selects clients, forms multiple client groups, and broadcasts the group assignments. Clients within each group coordinate their masks, apply the masks to their local model updates, and send the masked updates to the server. The server aggregates the client model updates within each group to form a candidate model and sends the candidate models to the clients for evaluation. The clients send their evaluation scores to the server, which likewise assigns rewards and penalties and performs robust aggregation based on the evaluation results. The server and clients obtain multiple plaintext candidate models. The sets of client updates aggregated into these models may overlap, making it possible to construct and solve a system of equations for the client updates. This risk can be avoided by limiting the size of MM or QQ, or by having the server check the sampled groups for privacy risks in advance. Our main focus is the effectiveness of the defense against Byzantine clients, so further constraints on privacy parameters are outside the scope of our study.

Dear Reviewer ffE7:

Thanks for your valuable comments. As the discussion phase is approaching its end, we would like to kindly confirm whether we have sufficiently addressed your concerns. Should there be any remaining questions requiring further clarification, please do not hesitate to let us know. If you are satisfied with our responses, we would greatly appreciate your consideration in adjusting the rating accordingly.

We sincerely look forward to your feedback.

Reviewer_ffE7OpenReview ↗

Official Comment

Thanks for the responses. After reading them, I decide to keep my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.