Home / Curated / Algorithms / RLR Robust Learning Rate Sign-Voting Backdoor Defense for Federated Learning

RLR Robust Learning Rate Sign-Voting Backdoor Defense for Federated Learning

Algorithms AIMech:RlrRobustLearningRateSignVotingBackdoorDefenseForFederatedLearning DRAFT

Aliases: Robust Learning Rate for Federated Learning, RLR Federated Backdoor Defense, FedAvg with Robust Learning Rate

A server-side federated learning backdoor defense that adjusts the server's learning rate vector eta_theta in {+eta, -eta}^d coordinate-wise at each communication round t according to the supermajority sign consensus of participating client updates Delta_t^k = w_t^k - w_t, setting eta_{theta, i} = +eta if |sum_{k in S_t} sgn(Delta_{t,i}^k)| >= theta and eta_{theta, i} = -eta otherwise, and updating the global model via element-wise product w_{t+1} = w_t + eta_theta odot g({Delta_t}) so that coordinates lacking honest sign consensus move in the loss-maximizing direction.

Identity Notes & Concept Scope Introduced by Ozdayi, Kantarcioglu, and Gel (AAAI 2021, Vol. 35, No. 10, pp. 9268-9276, DOI:10.1609/aaai.v35i10.17118, arXiv:2007.03767). Distinct from SignSGD with majority vote (AIMech:SignSgdAndMajorityVoteCompressedNonconvexOptimization, Bernstein et al. 2018), which discards update magnitudes and updates by eta * sgn(sum_k sgn(Delta_t^k)), whereas RLR preserves the full-precision aggregated update magnitude (e.g., FedAvg or coordinate-wise median) and flips the sign of the server learning rate to -eta on any coordinate where the absolute sum of update signs falls below threshold theta. Also distinct from FoolsGold (AIMech:FoolsGoldCosineSimilaritySybilDefenseForFederatedLearning), which assigns scalar per-client learning rates alpha_k based on historical cosine similarity rather than per-dimension sign-voting learning rates.

Claims & Grounded Evidence

4 Claims

Mechanism #c1-mechanism-coordinate-sign-vote-learning-rate-inversion

Given client updates {Delta_t^k}_{k in S_t} at round t and a server learning threshold hyperparameter theta, RLR sets the i-th coordinate server learning rate to eta_{theta, i} = +eta if |sum_{k in S_t} sgn(Delta_{t,i}^k)| >= theta and eta_{theta, i} = -eta otherwise (Eq. 5), equivalently multiplying the aggregated update by the diagonal unitary operator (2 I_t - I) with ||2 I_t - I||_2 = 1 so that coordinates below the sign-consensus threshold ascend rather than descend the empirical loss surface (Eqs. 6, 8-9).

supports Source: Defending against Backdoors in Federated Learning with Robust Learning Rate · Locator: Section 3 Equations (5)-(6) and Appendix D.1 Equations (8)-(9), pp. 3, 13

Scope #c2-scope-nonconvex-convergence-under-population-weighted-symmetry

Under L-Lipschitz smooth local objectives with bounded stochastic gradient variance sigma^2, population-weighted symmetric unimodal gradient noise (Eq. 7), universal update norm bound M, and sign-inversion probability Pr(1 - I_{|sum sgn(Delta_{t,i}^k)| >= theta} | F_t) <= p_0 < 0.25, Theorem 1 proves that RLR satisfies (1/T) sum_{t=0}^{T-1} E[||nabla f(w_hat_t)||^2] <= 2(f(w_hat_0) - f^*)/(eta T) + L^2 M^2 + L eta sigma^2 / n.

supports Source: Defending against Backdoors in Federated Learning with Robust Learning Rate · Locator: Section 3 Theorem 1 and Appendix D.2 Equations (10)-(21), pp. 3-4, 13-15

Application #c3-application-trojan-and-distributed-backdoor-defense

Combined with FedAvg, coordinate-wise median, or weight-clipping and Gaussian noise, RLR reduces trojan backdoor accuracy from 99-100% to 0-3.4% on i.i.d. Fashion-MNIST and non-i.i.d. Federated EMNIST and defends CIFAR-10 against partitioned Distributed Backdoor Attacks (DBA) where FoolsGold, Comed, and SignSGD fail.

supports Source: Defending against Backdoors in Federated Learning with Robust Learning Rate · Locator: Section 4, Tables 1-2, and Figures 2-5, pp. 4-7

Limitation #c4-limitation-threshold-selection-and-convergence-slowdown

Applying RLR from the beginning of training slows convergence and slightly reduces clean validation accuracy on non-i.i.d. client distributions where honest updates naturally disagree in sign, requiring careful tuning of theta in [K*F + 1, K - K*F] or delayed activation near convergence.

supports Source: Defending against Backdoors in Federated Learning with Robust Learning Rate · Locator: Section 4 (Removing Backdoor During Training), Section 6, and Appendix A & C.1 Figure 6, pp. 5, 7, 9-10

⚠️ Curated Limitations & Negative Results

Requires the number of corrupt agents in each round to be strictly smaller than the server sign-voting threshold theta.
Sign disagreement caused by severe non-i.i.d. data heterogeneity can trigger learning-rate sign flips on benign coordinates, slowing convergence and slightly reducing clean validation accuracy.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:FederatedLearning` Federated Learning	broad_match	RLR is a server-side coordinate sign-voting learning-rate defense against backdoor model poisoning in federated learning.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r529-primary-pdf`	Defending against Backdoors in Federated Learning with Robust Learning Rate Mustafa Safa Ozdayi, Murat Kantarcioglu, Yulia R. Gel	primary paper full_text	ID: https://dblp.org/rec/conf/aaai/OzdayiKG21 sha256:651d4101c1b3…
`r529-arxiv-abs`	Defending against Backdoors in Federated Learning with Robust Learning Rate (arXiv abstract) Mustafa Safa Ozdayi, Murat Kantarcioglu, Yulia R. Gel	primary paper abstract_only	ID: https://arxiv.org/abs/2007.03767 sha256:e4c723d41cc4…
`r529-metadata-cr`	Defending against Backdoors in Federated Learning with Robust Learning Rate (Crossref metadata) Mustafa Safa Ozdayi, Murat Kantarcioglu, Yulia R. Gel	primary paper abstract_only	ID: https://dblp.org/rec/conf/aaai/OzdayiKG21 sha256:63e62cbed269…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T10:12:00Z

Updated: 2026-10-19T10:12:00Z

Source File: data/curated/RlrRobustLearningRateSignVotingBackdoorDefenseForFederatedLearning.yaml · Schema: curated_concept.yaml

Explore across X-Mech

Related computational & biological Mechs: Embedding Map Network Graph ProteinTraitsMech (ESM-2 pLM) DUFMech (Domain pLM) TraitMech CultureMech All 14 X-Mech Repositories