acceptodds
Under review as a conference paper at ICLR 2027

When Trust Becomes a Liability: Bounding Persistent Influence under Temporally Adaptive Attacks in Federated Learning

Abstract

Reputation-based defenses in federated learning can make historical trust persist in aggregation. As a result, benign behavior in earlier rounds may continue to affect a client's aggregation weight after it turns malicious. We study this vulnerability under temporally adaptive adversaries and argue that aggregation influence should decay unless it is continually re-earned. We introduce DAIRE-FL (Decaying Aggregation Influence with Renewal and Enforcement), which separates behavioral acceptance, current aggregation weight, and future influence renewal to limit persistent historical influence. We derive finite-horizon guarantees that bound both inherited and re-earned influence after a behavioral pivot. Our threat model also considers pseudonymous re-enrollment and data reuse to distinguish persistent influence from increased adversarial participation. Experiments on CIFAR-10, CIFAR-100, FEMNIST, and Shakespeare under abrupt, gradual, threshold-aware, and layer-selective attacks show that DAIRE-FL substantially limits post-pivot adversarial influence while preserving benign-task performance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated learning, adaptive adversaries, reputation-based aggregation, model poisoning, trustworthy federated learning

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

This paper studies persistent aggregation influence in federated learning when previously benign clients become malicious after a pivot pp at round TpT_p. It proposes a method called DAIRE-FL, which maintains a decaying influence state, renews influence following accepted participation, and uses the pre-renewal state for current aggregation. Behavioral acceptance combines frozen and constrained adaptive conditional references, while attestation enforces renewal decisions. The paper derives pathwise finite-horizon bounds separating inherited influence from subsequent renewal. The paper performs experiments on CIFAR-10, CIFAR-100, FEMNIST, and Shakespeare under several adaptive attack configurations.

Critical strengths
  • The paper identifies a meaningful limitation of historical reputation. Benign participation before a behavioral pivot does not establish benign behavior afterward. Separating current aggregation authority from future renewal makes this concern explicit in practice.

  • The paper distinguishes pathwise bounds from conditional probabilistic refinements and acknowledges that enrollment does not establish identity scarcity, acceptance does not certify benign intent, and displacement bounds require bounded updates. These distinctions help prevent overly broad interpretations of the guarantees.

Critical weaknesses
  • The expected-contraction result requires malicious clients to obtain renewal infrequently enough, which is given in Eq. 13, βR<τmin⁡(1−e−λ)Va.\beta_R < \frac{\tau_{\min}(1-e^{-\lambda})}{V_a}. However, the paper does not derive a bound on the malicious acceptance probability from the proposed detector DAIRE-FL or its adaptation constraints described in Section 4.2. Therefore, the result explains what happens if malicious updates are rejected sufficiently often, but does not establish when the proposed detector ensures this condition against adaptive attackers. The pathwise bound in Theorem 1 remains valid without this assumption. Nevertheless, a theoretical connection between the detector and the required acceptance rate is missing, which limits the security conclusion that can be drawn from the analysis.

  • A client receives aggregation weight only when its update is accepted, and its normalized weight cannot exceed one. Therefore, after simple calculation, it shows

\quad\Longrightarrow\quad \Psi_k(H)\leq\mathcal R_k(H).$$ In simple terms, a client’s total aggregation weight cannot exceed the number of its accepted submissions. This bound holds independently of the proposed influence-decay mechanism in DAIRE-FL. For the stated default parameters, each of the three terms in the minimum in Theorem 1 is at least as large as this accepted-submission count, i.e., $\mathcal R_k(H)$. Thus, the theorem gives no tighter bound on cumulative aggregation weight in this setting. Although separating inherited and renewed influence is conceptually useful, but the authors should establish when their bound improves over the above inequality and what additional guarantee the decay mechanism provides. - Section 4 introduces influence decay, renewal, two detector references, PCA projections, threshold adaptation, enrollment proofs, and attestation in quick succession. Figure 1 is also dense, with small labels. A short example following one client through benign participation, malicious behavior, rejection, and re-entry would make the mechanism much easier to understand. Putting the method using a simple, concrete example will improve the paper.
Questions for authors
  • Under what conditions does the proposed detector ensure that malicious clients’ renewal rate satisfies Eq. (13)? Can this condition be established from the detector’s design and adaptation constraints, or does it require additional assumptions about the attacker?

  • Since pk(t)≤Rk(t)p_k(t)\leq R_k(t), we immediately obtain Ψk(H)≤Rk(H)\Psi_k(H)\leq\mathcal R_k(H). In which parameter regimes does Theorem 1 provide a strictly tighter bound? What additional guarantee does it offer at the paper’s default parameters?

  • The paper studies how influence changes over time, but Figure 2 mainly presents aggregate attack-success rates. Important timing details, including calibration and pivot rounds, are provided in the appendix. A timeline and a plot of client influence before and after the pivot would help readers understand what happens and why the defense works. It is a suggestion.

  • Section 5 states the mathematical conditions, but gives limited intuition about when the bound is informative. Add a clearer explanation of how inherited influence, subsequent renewals, and the horizon affect the guarantee, and what the bound establishes about the relationship between aggregation weight and actual damage in the model. In a nutshell, the practical meaning of the theorem needs more explanation.

5 comments
AuthorsOpenReview ↗

Responses to Reviewer NyQk

Weakness 1 (NyQk.W1): The expected-contraction result requires malicious clients to obtain renewal infrequently enough, which is given in Eq. (13). However, the paper does not derive a bound on the malicious acceptance probability from the proposed detector DAIRE-FL or its adaptation constraints described in Section 4.2. Therefore, the result explains what happens if malicious updates are rejected sufficiently often, but does not establish when the proposed detector ensures this condition against adaptive attackers. The pathwise bound in Theorem 1 remains valid without this assumption. Nevertheless, a theoretical connection between the detector and the required acceptance rate is missing, which limits the security conclusion that can be drawn from the analysis.

We thank the reviewer for identifying the acceptance–contraction gap. The frozen reference and constrained adaptation do not establish Eq. (13)'s history-uniform acceptance bound against adaptive attackers. They imply only a necessary condition: since rkt≥αEθL(xkt,Sk)r_k^t\ge\alpha E_{\theta^L}(x_k^t,S_k) and τt≤τL\tau_t\le\tau_L, acceptance requires EθL(xkt,Sk)≤τL/αE_{\theta^L}(x_k^t,S_k)\le\tau_L/\alpha. The ceiling limits admissible scores, not how often attackers satisfy it, including through benign recharge.

With conditionally uniform 1010-of-100100 sampling at the CIFAR defaults, a history-uniform conditional-acceptance bound below 0.28800.2880 suffices for Eq. (13)'s renewal threshold of 0.028800.02880. Neither the detector nor cohort-average rates establish this bound; sampling alone caps per-round renewal at 0.100.10, insufficient for contraction below τmin⁡\tau_{\min}. Contraction is therefore conditional, while Theorem 1 holds pathwise without an acceptance-probability assumption (Sections 4.2–5, pages 5–6; Appendix B.8, page 25).

Weakness 2 (NyQk.W2): A client receives aggregation weight only when its update is accepted, and its normalized weight cannot exceed one. In simple terms, a client's total aggregation weight cannot exceed the number of its accepted submissions. This bound holds independently of the proposed influence-decay mechanism in DAIRE-FL. For the stated default parameters, each of the three terms in the minimum in Theorem 1 is at least as large as this accepted-submission count. Thus, the theorem gives no tighter bound on cumulative aggregation weight in this setting. Although separating inherited and renewed influence is conceptually useful, the authors should establish when their bound improves over the above inequality and what additional guarantee the decay mechanism provides.

We thank the reviewer. At the CIFAR defaults, Eq. (14) does not improve Ψk(H)≤r\Psi_k(H)\le r, where r=Rk(H)r=\mathcal R_k(H). For r>0r>0, strict improvement requires max⁡{Zmin⁡,τmin⁡}(1−e−λ)>Va+bk/r\max\{Z_{\min},\tau_{\min}\}(1-e^{-\lambda})>V_a+b_k/r. The left side is at most 3.393.39 at the defaults and 4.9944.994 across listed positive-decay CIFAR settings, below Va=50.8V_a=50.8. An unevaluated analytical example gives 17.51<r=2017.51<r=20, although even continuous renewal approaches a balance below τmin⁡\tau_{\min} (Section 5, pages 6–7; Appendix D.8, page 33).

At these defaults, the timing-aware pathwise cap Ψk(H)≤Gk(H)≤r\Psi_k(H)\le G_k(H)\le r counts acceptances with eligible pre-renewal balance. Starting from 100100, balance falls below 3030 after 2525 renewal-free steps; the next acceptance restores future eligibility but has no current weight or detector-adaptation effect. Six spaced recharge–attack acceptances yield Gk=3G_k=3, versus Gk=6G_k=6 for six consecutive acceptances from the same pivot balance. This constrains when accepted updates contribute, not Eq. (14) uniformly or measured ASR (Section 5, page 7; Appendix B.7, pages 23–25).

Weakness 3 (NyQk.W3): Section 4 introduces influence decay, renewal, two detector references, PCA projections, threshold adaptation, enrollment proofs, and attestation in quick succession. Figure 1 is also dense, with small labels. A short example following one client through benign participation, malicious behavior, rejection, and re-entry would make the mechanism much easier to understand. Putting the method using a simple, concrete example will improve the paper.

We thank the reviewer for proposing a clearer Figure 1 and one-client walkthrough. We have simplified its labels while retaining enrollment, scoring, attestation, eligibility, and renewal. Section 4.1 explains re-entry, and Appendix B.7 provides the numerical Figure 3 (Section 4, page 4; Appendix B.7, pages 23–24). Benign renewals at rounds 54,64,74,84,9454,64,74,84,94 yield pivot balance Bi(100)≈100.50B_i(100)\approx100.50. If submitted post-pivot attacks are rejected, no renewals occur in rounds 100100–124124, reducing balance to Bi(125)≈28.79<30B_i(125)\approx28.79<30. An accepted benign recharge at round 125125 has no current weight or detector-adaptation effect but raises Bi(126)≈78.19B_i(126)\approx78.19; a later accepted attack can contribute. The constructed curve illustrates re-entry, not a measured training trace or permanent exclusion.

AuthorsOpenReview ↗

Responses to Reviewer NyQk’s Questions – Part 1

Question 1 (NyQk.Q1): Under what conditions does the proposed detector ensure that malicious clients' renewal rate satisfies Eq. (13)? Can this condition be established from the detector's design and adaptation constraints, or does it require additional assumptions about the attacker?

We thank the reviewer for asking whether the detector itself ensures Eq. (13). It does not. The frozen reference and constrained adaptation impose a necessary condition for acceptance, EθL(xkt,Sk)≤τL/αE_{\theta^L}(x_k^t,S_k)\le\tau_L/\alpha, but do not bound how often an adaptive attacker can submit updates satisfying it. A sufficient additional condition, for every pre-round history Gt\mathcal G_t with positive selection probability and every strategy under consideration, is Pr⁡[EθL(xkt,Sk)≤τL/α∣Jk(t)=1,Gt]≤βA\Pr[E_{\theta^L}(x_k^t,S_k)\le\tau_L/\alpha\mid J_k(t)=1,\mathcal G_t]\le\beta_A and πmax⁡βA<τmin⁡(1−e−λ)/Va\pi_{\max}\beta_A<\tau_{\min}(1-e^{-\lambda})/V_a. The first inequality bounds acceptance conditional on selection; the second places per-round renewal below Eq. (13)'s threshold. Under conditionally uniform 1010-of-100100 sampling at the controlled CIFAR defaults, a history-uniform βA<τmin⁡(1−e−λ)/(0.10Va)≈0.2880\beta_A<\tau_{\min}(1-e^{-\lambda})/(0.10V_a)\approx0.2880 would suffice.

The full threat model permits benign recharge by adversary-controlled identities, so neither the detector constraints nor cohort-average acceptance rates establish this condition for all allowed strategies. Selection alone caps per-round renewal at 0.100.10, above Eq. (13)'s approximately 0.028800.02880 threshold. Expected contraction is therefore conditional, while Theorem 1 remains pathwise without this assumption (Sections 4.2–5, pages 5–6; Appendix B.8, page 25).

Question 2 (NyQk.Q2): Since each normalized weight is at most one and is zero on rejection, we immediately obtain Ψk(H)≤Rk(H)\Psi_k(H)\le\mathcal R_k(H). In which parameter regimes does Theorem 1 provide a strictly tighter bound? What additional guarantee does it offer at the paper's default parameters?

We thank the reviewer for asking when Theorem 1 improves on acceptance counting and what it guarantees at the defaults. For r=Rk(H)>0r=\mathcal R_k(H)>0, Eq. (14) is strictly below rr exactly when max⁡{Zmin⁡,τmin⁡}(1−e−λ)>Va+bk/r\max\{Z_{\min},\tau_{\min}\}(1-e^{-\lambda})>V_a+b_k/r. This favors stronger decay and smaller renewal increments relative to the normalization and eligibility thresholds. The condition does not hold at the controlled CIFAR defaults or listed positive-decay settings with a positive normalizer; when all normalizers are zero, realized influence is zero. For λ=0.1\lambda=0.1, Va=1V_a=1, bk=τmin⁡=30b_k=\tau_{\min}=30, H=100H=100, and r=20r=20, Eq. (14) yields 17.51<2017.51<20. This demonstrates strictness in an unevaluated analytical regime with limited sustained eligibility (Section 5, pages 6–7; Appendix D.8, page 33).

At the unchanged CIFAR defaults, the additional guarantee comes from the separate timing-aware pathwise cap Ψk(H)≤Gk(H)≤r\Psi_k(H)\le G_k(H)\le r, which counts acceptances with eligible pre-renewal balance. Six consecutive acceptances yield Gk=6G_k=6, whereas six acceptances arranged as three spaced recharge–attack pairs yield Gk=3G_k=3. The three below-threshold recharge updates restore future eligibility but receive zero current weight and cannot influence detector adaptation (Section 5, page 7; Appendix B.7, pages 23–25).

We also provide an expected-weight refinement under conditionally uniform 1010-of-100100 sampling. Even if every selected update is accepted, the per-round renewal probability is at most 0.100.10. The Bernoulli envelope gives the following per-identity bounds for H=100H=100 without assuming independent attacker renewals (Appendix B.8, page 25):

Inherited balanceExpected count ceilingExpected weight ceiling
bk=100b_k=10010.0010.009.65959.6595
bk=Bmax⁡b_k=B_{\max} (warm-up ceiling)10.0010.009.96289.9628

These refinements quantify timing-dependent exclusion and expected influence under the evaluated parameters. Neither establishes an ASR bound or a uniformly tighter cap than realized acceptance counting.

AuthorsOpenReview ↗

Responses to Reviewer NyQk’s Questions – Part 2

Question 3 (NyQk.Q3): The paper studies how influence changes over time, but Figure 2 mainly presents aggregate attack-success rates. Important timing details, including calibration and pivot rounds, are provided in the appendix. A timeline and a plot of client influence before and after the pivot would help readers understand what happens and why the defense works. It is a suggestion.

We thank the reviewer for suggesting a timeline and a client-level view alongside the aggregate ASR results. We have retained Figure 1 for the method architecture, added a short re-entry example in Section 4.1, and added a separate balance trajectory as Figure 3 in Appendix B.7. For the illustrated controlled CIFAR-10 schedule, rounds 00–4949 are calibration, rounds 5050–9999 are benign online participation, and the pivot occurs at round 100100 (Section 6.1, page 8; Appendix C.2, page 27).

In the constructed trajectory, the client's pivot balance is approximately 100.50100.50. With no accepted renewals in rounds 100100–124124, it falls to 28.79<3028.79<30 before round 125125. An accepted recharge then has zero current weight and no detector-adaptation effect, but raises the next-round balance to approximately 78.1978.19, allowing a subsequent accepted update to contribute (Appendix B.7, Figure 3, pages 23–24). The plotted quantity is the pre-round balance, which determines eligibility; it is not a measured normalized-weight or ASR curve. Normalized weight also depends on the other clients' accepted updates.

Question 4 (NyQk.Q4): Section 5 states the mathematical conditions, but gives limited intuition about when the bound is informative. Add a clearer explanation of how inherited influence, subsequent renewals, and the horizon affect the guarantee, and what the bound establishes about the relationship between aggregation weight and actual damage in the model. In a nutshell, the practical meaning of the theorem needs more explanation.

We thank the reviewer for asking for the theorem's practical meaning. The inherited balance is the state available at the pivot; it decays in subsequent rounds. An accepted renewal increases only the next balance, not the weight of its own update. Thus a long gap can make an acceptance probationary: at the controlled CIFAR defaults, a balance of 100100 falls below the threshold of 3030 after 2525 rounds without renewal. Repeated accepted renewals can restore or maintain eligibility. Increasing the horizon dissipates the inherited component, but cumulative weight need not fall when renewals continue. At these defaults, Eq. (14) does not improve on counting acceptances; the separate timing-aware cap identifies acceptances that have zero weight because the pre-renewal balance is below threshold (Section 5, page 7; Appendices B.2 and B.7, pages 17–18 and 23–25).

Theorem 1 bounds normalized aggregation weight, not attack success or the difference between complete training trajectories. If the full update norm is separately bounded by ∣Δwk(t)∣≤δmax⁡\\|\Delta w_k(t)\\|\le\delta_{\max}, it also bounds the sum of direct weighted update magnitudes by Ωk(H)≤δmax⁡Ψk(H)\Omega_k(H)\le\delta_{\max}\Psi_k(H). Behavioral acceptance alone does not provide that norm cap, and conclusions about training outcomes require additional assumptions (Appendix B.6, pages 21–23).

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

This paper investigates how historical trust in federated learning can become a vulnerability when previously benign clients turn malicious. The proposed method, DAIRE-FL, maintains a decaying influence state for each client, decoupling update acceptance, current aggregation weight, and future influence renewal. This mechanism is combined with a dual-reference behavioral detector and an attestation protocol. The authors derive bounds on cumulative aggregation influence following a behavioral change and evaluate the method against abrupt, gradual, threshold-aware, and layer-selective attacks across four datasets.

Critical strengths
  • The distinction between accepting a client update and granting it aggregation weight is well-motivated. Specifically, the concept of "probationary" clients allows them to rebuild influence without immediately impacting the global model or the detector. This provides a concrete protocol contribution beyond merely replacing one detector with another, offering a structured way to handle changes in client behavior.

  • The experimental evaluation is reasonably thorough. The authors consider several forms of adaptive attacks, tune defense parameters without relying on attack outcomes, and report results over five seeds. The ablation study removing influence decay is particularly relevant: ASR increases from 12.6% to 29.8% on CIFAR-10, while malicious acceptance rates remain nearly unchanged. This supports the claim that managing historical influence is a distinct and important factor, independent of detecting malicious updates.

Critical weaknesses
  • Several reported influence statistics appear inconsistent with the paper's definitions. According to Eqs. (3), (4), and (67), the cumulative malicious aggregation weight (ΨM\Psi_M) and the mean honest aggregation mass (m^H\hat{m}_H) should satisfy ΨM(H)+Hm^H≤H\Psi_M(H) + H\hat{m}_H \le H, assuming they are measured over the same trajectories and post-pivot window. However, Table 12 reports ΨM=44.58\Psi_M=44.58 and m^H=0.642\hat{m}_H=0.642 for K=100K=100. With the default H=100H=100, their sum is 108.78108.78, which exceeds the maximum possible total weight of 100, right? Similar discrepancies are observed in other rows. Additionally, it is difficult to reconcile the renewal rates in Table 1 with the normalization diagnostics in Table 2. For instance, under the stated default cohort size and assuming matched evaluation runs, the abrupt-pivot renewal rate of 0.010 implies a diagnostic lower bound of approximately 149.9, rather than the reported 14.90. While these differences may stem from reporting conventions or separate runs, the paper does not provide sufficient detail to resolve these inconsistencies.

  • The current evidence does not sufficiently establish why the proposed influence mechanism is preferable to simpler alternatives. Theorem 1 bounds cumulative influence using inherited state and subsequent renewals. However, the aggregation rule inherently implies Ψk(H)≤Rk(H)\Psi_k(H) \le R_k(H), since a client can only receive weight when accepted, and that weight cannot exceed one. Under the default parameters, the bound in Eq. (14) is no tighter than this direct inequality. This suggests that the theoretical gain is less significant than the presentation implies. While the no-decay ablation demonstrates that decay helps, it does not prove that DAIRE-FL's specific renewal and weighting rules are necessary. A comparison with simpler baselines—such as using the same detector with uniform weights for accepted updates, or a conventional exponentially decayed reputation score—is missing. Without such comparisons, it is difficult to assess the added value of the relatively elaborate protocol.

Questions for authors
  1. Could you clarify how the influence quantities in Tables 1, 2, and 12 are aggregated? Do they use the same seeds, client cohorts, and post-pivot windows? Specifically, how do the results in Table 12 satisfy the normalization constraint implied by Eqs. (3)–(4)?

  2. Under the default configuration, is there any case where Eq. (14) provides a tighter bound than the trivial inequality Ψk(H)≤Rk(H)\Psi_k(H) \le R_k(H)? If not, what additional security property should readers take from Theorem 1?

  3. How would the results change if the same behavioral detector were used with uniform weighting of accepted clients, or with a simpler exponentially decayed reputation score?

  4. The threat model allows clients to re-enroll under new identities, but the experiments consider identities enrolled before calibration. Since a newly enrolled identity starts with B0=100B_0=100 (above τmin⁡=30\tau_{\min}=30), could an attacker repeatedly replace identities to avoid the consequences of influence decay?

3 comments
AuthorsOpenReview ↗

Responses to Reviewer RKVx

Weakness 1 (RKVx.W1): Several reported influence statistics appear inconsistent with the paper's definitions. According to Eqs. (3), (4), and (67), the cumulative malicious aggregation weight (ΨM\Psi_M) and the mean honest aggregation mass (m^H\widehat{m}_H) should satisfy ΨM(H)+Hm^H≤H\Psi_M(H) + H\widehat{m}_H \le H, assuming they are measured over the same trajectories and post-pivot window. However, Table 12 reports ΨM=44.58\Psi_M = 44.58 and m^H=0.642\widehat{m}_H = 0.642 for K=100K = 100. With the default H=100H = 100, their sum is 108.78, which exceeds the maximum possible total weight of 100, right? Similar discrepancies are observed in other rows. Additionally, it is difficult to reconcile the renewal rates in Table 1 with the normalization diagnostics in Table 2. For instance, under the stated default cohort size and assuming matched evaluation runs, the abrupt-pivot renewal rate of 0.010 implies a diagnostic lower bound of approximately 149.9, rather than the reported 14.90. While these differences may stem from reporting conventions or separate runs, the paper does not provide sufficient detail to resolve these inconsistencies.

We thank the reviewer for these checks. For exhaustive cohorts on a common post-pivot trace, Hm^H+ΨM(H)=H−n0≤HH\widehat m_{\mathrm H}+\Psi_{\mathrm M}(H)=H-n_0\le H (n0n_0: zero-normalizer rounds); matched-seed averages preserve this identity. Yet original Table 12 implies 108.78108.78 (K=100K=100) and 101.34101.34 (K=10K=10) for H=100H=100; rounding cannot explain either excess. The Table 1/2 renewal inference requires matched seeds, cohorts, and windows. If matched, 10(100)(0.010)=1010(100)(0.010)=10 mean renewals and Zmin⁡obs≤10log⁡(1+Bmax⁡)≈69.49Z_{\min}^{\mathrm{obs}}\le10\log(1+B_{\max})\approx69.49 imply a diagnostic of at least 149.9149.9, not 14.9014.90. Independently, original Table 2's within-trace weight-to-diagnostic ratios (0.380.38–0.400.40) exceed log⁡(31)/30≈0.11447\log(31)/30\approx0.11447, regardless of cross-table pairing. We have withdrawn original Tables 2 and 12, proliferation Figure 3, and the associated validation claim. Section 6.2.2 clarifies acceptance, renewal, and eligibility; Appendix D.2 retains trace-consistency checks. The pathwise theorem remains independent.

Weakness 2 (RKVx.W2): The current evidence does not sufficiently establish why the proposed influence mechanism is preferable to simpler alternatives. Theorem 1 bounds cumulative influence using inherited state and subsequent renewals. However, the aggregation rule inherently implies Ψk(H)≤Rk(H)\Psi_k(H)\le\mathcal R_k(H), since a client can only receive weight when accepted, and that weight cannot exceed one. Under the default parameters, the bound in Eq. (14) is no tighter than this direct inequality. This suggests that the theoretical gain is less significant than the presentation implies. While the no-decay ablation demonstrates that decay helps, it does not prove that DAIRE-FL's specific renewal and weighting rules are necessary. A comparison with simpler baselines—such as using the same detector with uniform weights for accepted updates, or a conventional exponentially decayed reputation score—is missing. Without such comparisons, it is difficult to assess the added value of the relatively elaborate protocol.

We thank the reviewer. Eq. (14) does not improve on Ψk(H)≤Rk(H)\Psi_k(H)\le\mathcal R_k(H) at the CIFAR defaults, and the no-decay ablation tests forgetting, not weighting-rule necessity. While the balance is a rescaled EMA of renewals, DAIRE-FL uses pre-renewal eligibility to gate aggregation and detector adaptation. Its timing-aware cap gives Gk=3G_k=3 for six accepted updates in the spaced schedule (Section 5, page 7; Appendix B.7, pages 23–25).

We have added a five-paired-seed CIFAR-10 Mode A control sharing the detector, renewal rule, and detector-adaptation gate; only aggregation differs. DAIRE-FL reduces ASR from 11.7±1.8%11.7\pm1.8\% (uniform) to 7.0±1.8%7.0\pm1.8\% (paired improvement 4.7±1.34.7\pm1.3 points); clean accuracy changes from 82.0±0.6%82.0\pm0.6\% to 81.8±0.8%81.8\pm0.8\% (paired difference −0.2±0.2-0.2\pm0.2 points). This supports the aggregation rule in Mode A, without attributing the gain to gating versus the log map (Section 6.2.4, Table 3, page 10; Appendix C.3, page 28). Appendix B.7 fixes acceptances and balances with one honest rival:

Weighting ruleAll six acceptancesThree attack acceptances
Uniform over accepted updates3.00003.00001.50001.5000
Ungated linear EMA0.32830.32830.23550.2355
Ungated logarithmic weighting2.17052.17051.17141.1714
DAIRE-FL1.17141.17141.17141.1714

Against the same ungated logarithmic map, gating eliminates recharge-round weight, reducing focal cumulative weight from 2.17052.1705 to 1.17141.1714, without reducing the three attack-round weights. Ungated linear EMA assigns less weight here by preserving larger balance differences. Thus the replay isolates the gate's immediate effect, not superiority over EMA or a training ASR ranking; a detector-matched EMA training control was not performed.

AuthorsOpenReview ↗

Responses to Reviewer RKVx’s Questions – Part 1

Question 1 (RKVx.Q1): Could you clarify how the influence quantities in Tables 1, 2, and 12 are aggregated? Do they use the same seeds, client cohorts, and post-pivot windows? Specifically, how do the results in Table 12 satisfy the normalization constraint implied by Eqs. (3)–(4)?

We thank the reviewer for requesting clarification. The original submission specified five seeds and post-pivot rounds 100100–199199 (H=100H=100). Tables 1 and 2 concerned the main cohort of 90 honest and 10 adversarial identities, whereas Table 12 used 100 honest identities plus KK adversarial identities. The nominal horizon was unchanged, but exact seed-wise alignment between Tables 1 and 2 cannot be established from their aggregates; hence the cross-table diagnostic estimate of approximately 149.9149.9 is conditional on matched traces.

Conditional acceptance and renewal use the same adversarial acceptance count, divided by adversarial submissions and KHKH, respectively. Malicious influence is cumulative, whereas honest aggregation mass is averaged over HH. For an exhaustive cohort on a common trace, Eqs. (3)–(4) require Hm^H+ΨM(H)=H−n0≤HH\widehat m_{\mathrm H}+\Psi_{\mathrm M}(H)=H-n_0\le H, where n0n_0 counts zero-normalizer rounds. This identity also holds after averaging matched seeds. The original Table 12 implied 101.34101.34 for K=10K=10 and 108.78108.78 for K=100K=100, exceeding H=100H=100 and thus violating the stated normalization constraint. We have therefore withdrawn Tables 2 and 12 and proliferation Figure 3 of the original submission, along with their empirical correspondence claim. The revised manuscript retains the benchmark comparisons, clarifies acceptance and influence eligibility in Section 6.2.2, and provides trace-level consistency conditions in Appendix D.2. Theorem 1 remains valid independently of the withdrawn statistics.

Question 2 (RKVx.Q2): Under the default configuration, is there any case where Eq. (14) provides a tighter bound than the trivial inequality Ψk(H)≤Rk(H)\Psi_k(H)\le\mathcal R_k(H)? If not, what additional security property should readers take from Theorem 1?

We thank the reviewer for asking whether Eq. (14) improves on acceptance counting at the defaults. No. For r=Rk(H)>0r=\mathcal R_k(H)>0 and a positive normalizer, it is strictly below rr only if max⁡Zmin⁡,τmin⁡(1−e−λ)>Va+bk/r\max\\{Z_{\min},\tau_{\min}\\}(1-e^{-\lambda})>V_a+b_k/r. At the controlled CIFAR defaults, the left side is at most about 3.393.39, below Va=50.8V_a=50.8, so this condition cannot hold. If r=0r=0 or every normalizer is zero, realized weight is zero directly. An initially eligible client that is the sole positive-weight contributor on its accepted rounds can attain Ψk(H)=r\Psi_k(H)=r (Section 5, page 7).

Theorem 1 provides pathwise accounting of inherited balance and subsequent renewals without assuming a detector acceptance rate; we do not present Eq. (14) as a tighter default-setting cap. The additional guarantee at those settings comes from the separate timing-aware cap Ψk(H)≤Gk(H)≤r\Psi_k(H)\le G_k(H)\le r. From the same pivot balance, six consecutive acceptances give Gk=6G_k=6, whereas six acceptances arranged as three spaced recharge–attack pairs give Gk=3G_k=3: each below-threshold recharge renews future balance but has zero current aggregation weight and no detector-adaptation effect. This is a trace-specific guarantee about when acceptance can contribute, not a bound on ASR or on later accepted attacks (Section 5, page 7; Appendix B.7, pages 23–25).

AuthorsOpenReview ↗

Responses to Reviewer RKVx’s Questions – Part 2

Question 3 (RKVx.Q3): How would the results change if the same behavioral detector were used with uniform weighting of accepted clients, or with a simpler exponentially decayed reputation score?

We thank the reviewer for proposing uniform aggregation and an exponentially decayed reputation score as controls. We have added a five-paired-seed CIFAR-10 Mode A training comparison with uniform aggregation of accepted updates. Both arms retain the same detector rules, renewal rule, and pre-renewal detector-adaptation gate; only model aggregation changes. DAIRE-FL achieves ASR 7.0±1.87.0\pm1.8\\% versus 11.7±1.811.7\pm1.8\\% with uniform aggregation, a paired difference of −4.7±1.3-4.7\pm1.3 percentage points. Clean accuracy is 81.8±0.881.8\pm0.8\\% versus 82.0±0.682.0\pm0.6\\%, a paired difference of −0.2±0.2-0.2\pm0.2 points (Section 6.2.4, Table 3, page 10; Appendix C.3, page 28).

We do not report a detector-matched EMA training run. Indeed, with a=e−λa=e^{-\lambda} and si(t)=(1−a)Bi(t)/Vas_i(t)=(1-a)B_i(t)/V_a, the balance recurrence is already a rescaled EMA: si(t+1)=asi(t)+(1−a)Ri(t)s_i(t+1)=as_i(t)+(1-a)R_i(t). The distinction is how the pre-renewal state gates current use. In the fixed-acceptance analytical replay, cumulative focal weight is 1.17141.1714 under DAIRE-FL, 3.00003.0000 under uniform weighting, and 0.32830.3283 under ungated linear EMA; attack-round weight is 1.17141.1714, 1.50001.5000, and 0.23550.2355, respectively (Appendix B.7, pages 23–24). EMA assigns less weight than DAIRE-FL on this trace, so we claim no dominance over it. These replay weights do not predict an EMA training ASR or clean-accuracy result.

Question 4 (RKVx.Q4): The threat model allows clients to re-enroll under new identities, but the experiments consider identities enrolled before calibration. Since a newly enrolled identity starts with B0=100B_0=100 (above τmin⁡=30\tau_{\min}=30), could an attacker repeatedly replace identities to avoid the consequences of influence decay?

We thank the reviewer for highlighting identity replacement as a way to reset influence. Yes: at the controlled CIFAR defaults, a fresh pseudonym starts at B0=100>τmin⁡=30B_0=100>\tau_{\min}=30 rather than inheriting its predecessor's depleted balance. Its first selected and accepted update can therefore receive weight while it remains eligible. Enrollment, behavioral acceptance, and quorum attestations are still required, but they neither link pseudonyms nor establish identity scarcity (Section 3.3, page 3; Section 4.1, page 4; Appendix A.1, Table 4, page 14).

Theorem 1 bounds influence per identity, and the expected-weight sampling refinement additionally assumes a fixed pool. Neither bounds one actor's total influence across replacement identities. The retained empirical evaluation uses identities enrolled before calibration. The original static-proliferation experiment, whose results were withdrawn, varied pre-enrolled identities rather than testing post-pivot identity replacement. We have made this scope limit explicit and do not claim robustness against dynamic identity rotation (Section 3.3, page 3; Appendix D.5, page 32).

Reviewer_ySRtOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies how historical trust can remain useful to malicious clients after they change their behavior in federated learning. The proposed DAIRE-FL gradually reduces each client’s aggregation influence and replenishes it through accepted participation. It separates current aggregation from future renewal and constrains how the behavioral detector adapts. The paper provides finite-horizon influence bounds and evaluates the framework across several datasets and adaptive attacks

Critical strengths
  1. the problem is meaningful and well motivated for the change of behavior for users. 2.The separation between current aggragation and future renewal is reasonable which allows the client to recover the eligibility
Critical weaknesses

1.The added value of maintaining historical influence is not fully isolated. The paper includes a useful ablation without decay, but this still retains historical weighting. A comparison using the same detector followed by uniform aggregation of accepted updates would help establish whether the proposed influence mechanism offers benefits over a simpler filtering approach.

  1. Not sure if theoretical bound is practical and clear. Under the reported CIFAR configuration, the main bound is weaker than simply counting accepted submissions, since each submission can contribute at most one unit of normalized weight. The analysis separates inherited and renewed influence, but it remains unclear what additional security insight the bound provides in the evaluated setting.

  2. Some reported influence statistics appear inconsistent with their definitions. The diagnostic values in Table 2 do not match the constraints implied by the stated parameters and Appendix D.2. Table 12 also implies that honest and malicious aggregation weights together exceed the available total under the default evaluation horizon. These discrepancies need clarification before the experiments can support the claimed correspondence with theory.

Questions for authors

1.How does DAIRE-FL compare with the same behavioral detector followed by uniform aggregation of accepted updates? How much benefit comes specifically from maintaining and renewing historical influence?

2.In which practical settings is the proposed bound tighter than the simple bound based on accepted submissions? Can this improvement be demonstrated using the experimental configurations?

3.Can you clarify the parameters, evaluation windows, and calculations used in Tables 2 and 12? In particular, why do the reported honest and malicious weights in Table 12 sum to more than 100% under the default horizon?

4.How does the method behave when an attacker repeatedly alternates between benign participation to rebuild influence and malicious updates to exploit it? This seems particularly relevant to evaluating the proposed renewal mechanism.

5 comments
AuthorsOpenReview ↗

Responses to Reviewer ySRt

Weakness 1 (ySRt.W1): The added value of maintaining historical influence is not fully isolated. The paper includes a useful ablation without decay, but this still retains historical weighting. A comparison using the same detector followed by uniform aggregation of accepted updates would help establish whether the proposed influence mechanism offers benefits over a simpler filtering approach.

We thank the reviewer for noting that the no-decay ablation retains historical weighting and for requesting a detector-matched uniform control. We have added a five-paired-seed CIFAR-10 Mode A training comparison in which both arms share the behavioral detector, renewal rule, and pre-renewal detector-adaptation gate; only model aggregation differs. The uniform arm assigns equal weight to all accepted updates, including probationary ones. Results are summarized below.

CIFAR-10 Mode A; mean ±\pm sample SD over five paired seeds.

Aggregation ruleASR (%)Clean accuracy (%)
DAIRE-FL (proposed)7.0±1.87.0\pm1.881.8±0.881.8\pm0.8
Same detector + uniform aggregation11.7±1.811.7\pm1.882.0±0.682.0\pm0.6
Paired difference (DAIRE-FL minus uniform, points)−4.7±1.3-4.7\pm1.3−0.2±0.2-0.2\pm0.2

DAIRE-FL lowers mean ASR by 4.74.7 points with a 0.20.2-point reduction in clean accuracy (Section 6.2.4, Table 3, page 10). These within-seed differences come from a separate paired evaluation, not the Table 1 aggregate. The result supports DAIRE-FL's aggregation rule in Mode A but does not isolate gating from logarithmic weighting. Under fixed acceptances and an unchanged logarithmic map, pre-renewal gating reduces cumulative focal weight from 2.17052.1705 to 1.17141.1714 (Appendix B.7, pages 23–25). This isolates the gate's immediate weight effect, not its contribution to training ASR.

Weakness 2 (ySRt.W2): Not sure if theoretical bound is practical and clear. Under the reported CIFAR configuration, the main bound is weaker than simply counting accepted submissions, since each submission can contribute at most one unit of normalized weight. The analysis separates inherited and renewed influence, but it remains unclear what additional security insight the bound provides in the evaluated setting.

We thank the reviewer for this observation. We agree that in controlled CIFAR defaults, Eq. (14) does not improve the direct cap Ψk(H)≤r\Psi_k(H)\le r, where r=Rk(H)r=\mathcal R_k(H). For both six-acceptance schedules below, it returns the horizon cap H=100H=100. We have therefore clarified that its inherited-balance and renewal decomposition is not a tighter default-setting bound (Section 5, pages 6–7). The additional security insight comes from the separate timing-aware cap Ψk(H)≤Gk(H)≤r\Psi_k(H)\le G_k(H)\le r. It counts acceptances whose pre-renewal balance meets the eligibility threshold and can be computed from the pivot balance and acceptance times. In the worked trajectory, bk≈100.50b_k\approx100.50 decays without renewal to Bk(125)≈28.79<30B_k(125)\approx28.79<30. Acceptance at round 125125 restores future eligibility but has zero current weight (Appendix B.7, Figure 3, pages 23–24).

Acceptance scheduleEq. (14)Count cap rrTiming cap Gk(H)G_k(H)
Six consecutive acceptances, rounds 100100–1051051001006666
Three recharge–attack pairs: (125,126)(125,126), (156,157)(156,157), (187,188)(187,188)1001006633

At the unchanged CIFAR parameters, the spaced schedule allows at most three contributing rounds from six acceptances. This timing-specific pathwise guarantee establishes neither training frequency nor an ASR bound (Section 5, page 7; Appendix B.7, pages 23–24).

Weakness 3 (ySRt.W3): Some reported influence statistics appear inconsistent with their definitions. The diagnostic values in Table 2 do not match the constraints implied by the stated parameters and Appendix D.2. Table 12 also implies that honest and malicious aggregation weights together exceed the available total under the default evaluation horizon. These discrepancies need clarification before the experiments can support the claimed correspondence with theory.

We thank the reviewer for pointing out the inconsistencies. For exhaustive cohorts on one trace, normalization requires Hm^H+ΨM(H)=H−n0≤HH\widehat m_{\mathrm H}+\Psi_{\mathrm M}(H)=H-n_0\le H, where n0n_0 counts zero-normalizer rounds; matched-seed averages preserve this identity. Yet Table 12 implies 108.78108.78 (K=100K=100) and 101.34101.34 (K=10K=10) for H=100H=100. Independently, Table 2's within-trace weight-to-diagnostic ratios (0.380.38--0.400.40) exceed log⁡(31)/30≈0.11447\log(31)/30\approx0.11447 at τmin⁡=30\tau_{\min}=30. Neither check requires pairing Tables 1 and 2, and rounding cannot explain either. We have withdrawn original Tables 2 and 12, proliferation Figure 3, and their empirical-validation claim. Section 6.2.2 clarifies acceptance, renewal, and eligibility; Appendix D.2 retains trace-consistency checks. The pathwise theorem remains independent of the withdrawn results.

AuthorsOpenReview ↗

Responses to Reviewer ySRt’s Questions – Part 1

Question 1 (ySRt.Q1): How does DAIRE-FL compare with the same behavioral detector followed by uniform aggregation of accepted updates? How much benefit comes specifically from maintaining and renewing historical influence?

We thank the reviewer for requesting a detector-matched uniform-aggregation comparison. We have added a five-paired-seed CIFAR-10 Mode A training control. Both arms share the detector, renewal rule, and pre-renewal adaptation gate; only aggregation differs. DAIRE-FL achieves ASR 7.0±1.8%7.0\pm1.8\% versus 11.7±1.8%11.7\pm1.8\% with uniform aggregation, a within-seed difference of −4.7±1.3-4.7\pm1.3 percentage points. Clean accuracy is 81.8±0.8%81.8\pm0.8\% versus 82.0±0.6%82.0\pm0.6\%, a paired difference of −0.2±0.2-0.2\pm0.2 points (Section 6.2.4, Table 3, page 10; Appendix C.3, page 28). This supports DAIRE-FL's aggregation rule in Mode A but does not isolate gating from logarithmic weighting.

We separately isolate the gate's immediate weight effect at fixed acceptances and pre-round balances. In the six-acceptance analytical replay, the focal client's cumulative normalized weight is 3.00003.0000 with uniform aggregation, 2.17052.1705 with the logarithmic map but no gate, and 1.17141.1714 with DAIRE-FL. The gate removes weight from accepted recharge rounds, but the three later accepted attacks retain cumulative weight 1.17141.1714 under both logarithmic variants (Appendix B.7, pages 23–25). These are controlled weight calculations, not a separate end-to-end ASR attribution.

Question 2 (ySRt.Q2): In which practical settings is the proposed bound tighter than the simple bound based on accepted submissions? Can this improvement be demonstrated using the experimental configurations?

We thank the reviewer for asking when the bound improves on acceptance counting and whether this occurs at the evaluated settings. For r=Rk(H)>0r=\mathcal R_k(H)>0, Eq. (14) is strictly below rr exactly when max⁡{Zmin⁡,τmin⁡}(1−e−λ)>Va+bk/r\max\{Z_{\min},\tau_{\min}\}(1-e^{-\lambda})>V_a+b_k/r. This favors faster decay or larger normalizer bounds relative to renewal and inherited balance. However, across the listed positive-decay CIFAR settings, the left side is at most 4.9944.994, while the right side is at least 50.850.8. Thus Eq. (14) does not improve on counting at those settings when a positive normalizer exists; if all normalizers are zero, realized weight is zero directly (Section 5, page 7; Appendix D.8, page 33).

At the unchanged CIFAR defaults, the separate timing-aware cap can be stricter. For the constructed six-acceptance schedule with recharge–attack pairs (125,126)(125,126), (156,157)(156,157), and (187,188)(187,188), three recharge updates are accepted below the pre-renewal eligibility threshold, giving Ψk(H)≤Gk(H)=3<r=6\Psi_k(H)\le G_k(H)=3<r=6. This demonstrates a timing-dependent improvement at the experimental parameter values, not an improvement of Eq. (14) or a frequency or ASR result measured from training traces (Section 5, page 7; Appendix B.7, pages 23–24).

Question 3 (ySRt.Q3): Can you clarify the parameters, evaluation windows, and calculations used in Tables 2 and 12? In particular, why do the reported honest and malicious weights in Table 12 sum to more than 100% under the default horizon?

We thank the reviewer. Under the stated definitions, normalized weights cannot exceed the available total: on a common HH-round trace with exhaustive honest and adversarial cohorts, m^H+ΨM(H)/H=1−n0/H≤1\widehat m_{\mathrm H}+\Psi_{\mathrm M}(H)/H=1-n_0/H\le1, where n0n_0 counts zero-normalizer rounds. The honest entry is a per-round mean, while malicious weight is cumulative. With H=100H=100, the K=10K=10 row gives 0.944+6.94/100=1.01340.944+6.94/100=1.0134, and the K=100K=100 row gives 0.642+44.58/100=1.08780.642+44.58/100=1.0878. Thus the excess remains after putting both quantities on the same scale. These values violate the normalization identity under the stated common-window definitions and cannot represent additional aggregation mass.

The stated post-pivot window is rounds 100100–199199 in a 200-round run, with five evaluation seeds and n=10n=10 sampled identities per round. The main pool contains 90 honest and 10 adversarial identities. The submitted proliferation sweep instead added KK adversarial identities to 100 honest identities; its K=10K=10 row therefore had 110 enrolled identities, but still at most one unit of normalized weight per round. Controlled CIFAR uses B0=100B_0=100, λ=0.05\lambda=0.05, Va=50.8V_a=50.8, and τmin⁡=30\tau_{\min}=30; VaV_a is calibrated separately for other datasets (Section 6.1, page 8; Appendix C.2, Table 5, page 27). Submitted Table 2 used an ex post diagnostic computed per trace: summed adversarial pivot balances plus VaV_a times their accepted-renewal count, divided by Zmin⁡obs(1−e−λ)Z_{\min}^{\mathrm{obs}}(1-e^{-\lambda}), where Zmin⁡obsZ_{\min}^{\mathrm{obs}} is the trace's smallest positive normalizer. The affected influence summaries have been withdrawn from the manuscript. Appendix D.2 now presents the general trace-level consistency conditions, including normalization, diagnostic ratios, and renewal accounting.

AuthorsOpenReview ↗

Responses to Reviewer ySRt’s Questions – Part 2

Question 4 (ySRt.Q4): How does the method behave when an attacker repeatedly alternates between benign participation to rebuild influence and malicious updates to exploit it? This seems particularly relevant to evaluating the proposed renewal mechanism.

We thank the reviewer for asking how DAIRE-FL handles an attacker who alternates between benign updates to rebuild influence and malicious updates to use it. An accepted update below τmin⁡\tau_{\min} renews future balance but has zero current model weight and no detector-adaptation effect. At the controlled CIFAR defaults, one accepted recharge restores eligibility even from zero because Va=50.8>τmin⁡=30V_a=50.8>\tau_{\min}=30. A later malicious update can therefore receive weight if it passes the detector. The gate blocks the recharge round, not subsequent exploitation (Section 4.1, page 4). Appendix B.7 (pages 23–24) follows three stipulated accepted recharge–attack pairs, with no other focal renewals. One honest rival starts with the same pivot balance and is accepted every post-pivot round. The resulting pre-round balances and normalized weights are:

Recharge / attack roundsBalance before rechargeBalance before attackDAIRE-FL weight: recharge / attack
125/126125/12628.7928.7978.1978.190/0.39600/0.3960
156/157156/15729.3629.3678.7378.730/0.38840/0.3884
187/188187/18829.4829.4878.8578.850/0.38700/0.3870

The gate lowers the focal identity's total weight from 2.17052.1705 under the same ungated logarithmic map to 1.17141.1714, but weight on the three later accepted attacks remains 1.17141.1714 in both. If renewals keep the balance eligible, even the benign rounds need not have zero weight. This is a fixed-acceptance analytical replay, not a measured ASR result or an end-to-end training test of a balance-aware on–off attacker (Section 6.2.4, pages 9–10; Appendix B.7, pages 23–25).

Comments on the paper

1 comment

Dear reviewers, as the deadline is approaching, please check the rebuttal and consider participating in the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.