---
title: 'FedRansel: Heterogeneous Federated Selection'
url: https://www.emergentmind.com/topics/fedransel
type: topic
---

# FedRansel: Heterogeneous Federated Selection

Searching arXiv for the exact term and related papers to ground the article in the cited literature.
FedRansel is an overloaded term in recent federated-learning literature rather than a single standardized algorithm. In the cited arXiv record, it denotes standard random client selection in contribution valuation, the randomized LoRA subparameter selection mechanism inside FedRand, secure federated retrieval and answer selection across private data silos, a misspelling or alternative naming of FedeRank, and a pseudo-centralized random-intersection aggregation layer for quantum-enhanced fraud detection [2602.05693] [2503.07216] [2603.25374] [2012.11328] [2507.22908]. This suggests that the term must be interpreted from the surrounding problem formulation rather than from the label alone.

## 1. Terminological status

The usages that appear in the cited literature are heterogeneous and technically non-equivalent.

| Usage context | Meaning of “FedRansel” |
| --- | --- |
| Contribution valuation in federated learning | Standard random client selection per round; if this is the only randomness, this is FedRansel |
| LoRA-based federated VLM fine-tuning | Randomized LoRA subparameter selection inside FedRand: upload either $A$ or $B$ and keep the complementary family private |
| Federated retrieval-augmented generation | Secure federated retrieval and answer selection across private data silos |
| Federated recommendation | A misspelling or alternative naming of FedeRank |
| Financial fraud detection | Random parameter sampling at clients plus intersection-based robust averaging at the server |

Taken together, these usages indicate that the shared lexical motif is federated selection, but the selected object differs by setting: clients, parameter families, retrieved evidence, feedback signals, or parameter coordinates.

## 2. FedRansel as random client selection in contribution valuation

In "FedRandom: Sampling Consistent and Accurate Contribution Values in Federated Learning" [2602.05693], FedRansel is not the proposed mitigation for valuation instability. The paper formulates participant valuation through a coalition utility $v(S)$ over client subsets $S \subseteq N$, and recalls the canonical Shapley value
$$
\phi_i = \sum_{S \subseteq N \setminus \{i\}} \frac{|S|!(n-|S|-1)!}{n!}\big[v(S \cup \{i\}) - v(S)\big].
$$
Contribution assessment is performed round-by-round with Multi-Round Reconstruction Shapley values, while empirical bias is measured relative to the size-based reference baseline $p_i = D_i / \sum_j D_j$.

Within the algorithmic description, standard FL client sampling remains available at each round through the selection of a subset $C_t$; the paper states that this is FedRansel if this is the only randomness. FedRandom introduces a distinct stochastic mechanism: instead of randomizing only participation, it randomizes the server aggregation rule at every round, with $a_t \sim \mathrm{Unif}(S)$ over $S=\{\mathrm{FedAvg}, \mathrm{FedAvgM}, \mathrm{FedAdagrad}, \mathrm{FedAdam}, \mathrm{FedYogi}\}$. The distinction is therefore structural. FedRansel randomizes which clients participate, whereas FedRandom randomizes how their updates are aggregated.

The paper also makes the difference in purpose explicit. FedRansel is described as targeting participation fairness and system efficiency, while FedRandom targets valuation robustness. The reported empirical gains—reducing the overall distance to the ground truth by more than a third in half of all evaluated scenarios, improving stability in more than 90% of cases, and lowering variance in 303 of 324 scenarios—belong to FedRandom rather than to FedRansel.

## 3. FedRansel as randomized LoRA subparameter selection

In "FedRand: Enhancing Privacy in Federated Learning with Randomized LoRA Subparameter Updates" [2503.07216], FedRansel is understood as the randomized selection mechanism at the heart of FedRand. The model uses Low-Rank Adaptation, with a frozen base weight $W_0$ and a low-rank update $\Delta W$, written in the paper’s notation as
$$
W^{(k,l)}_{r,t} = W^{(l)}_0 + A^{(k,l)}_{r,t} B^{(k,l)}_{r,t}.
$$
At each round, each participating client randomly chooses one LoRA subparameter family to be public: either all $A$ matrices across layers or all $B$ matrices across layers. The choice is driven by
$$
u^{(k)} \sim \mathrm{Uniform}(0,1), \qquad a_k = 1_{u^{(k)} < \rho},
$$
with default $\rho=0.5$. If $a_k=1$, the client uploads $A$ and keeps $B$ private; otherwise it uploads $B$ and keeps $A$ private.

The server aggregates only the family actually received. With client weights $w_k=n_k/m_r$, the normalization factors are
$$
\alpha = \sum_{k\in S_r} w_k \cdot 1_{a_k=1}, \qquad
\beta = \sum_{k\in S_r} w_k \cdot 1_{a_k\neq 1},
$$
and each family is averaged only over the subset of clients that uploaded it. This normalization is central to the protocol: the ablation without normalization causes marked accuracy drops.

The privacy rationale is architectural rather than formal. Because one LoRA family is never uploaded in a given round, the server cannot exactly reconstruct a client’s effective adapter, and the paper argues that this reduces the mutual information between the server’s view and client-specific training signals. The evaluated attack is a server-side membership inference attack using MaxRényi-K% scores and AUROC, with lower AUROC indicating stronger defense. Empirically, FedRand improves robustness against MIAs while maintaining accuracy comparable to full sharing. On MSCOCO server-side performance, BLEU-4 is reported as $\approx 31.89$ for FedRand versus $31.80$ for FedAvg, ROUGE as $\approx 55.15$ versus $55.29$, and CIDEr as $\approx 110.27$ versus $111.08$; on ScienceQA, FedRand achieves $80.12\%$ versus $81.50\%$ for FedAvg. For privacy robustness, MaxRényi-10% on captions for the server model is $66.61$ for FedRand versus $70.22$ for FedAvg. Communication per round is reduced from $2L$ in full LoRA sharing to $1.5L$, a reduction of $\approx 25\%$.

## 4. FedRansel as secure federated retrieval and answer selection

In "Supercharging Federated Intelligence Retrieval" [2603.25374], FedRansel is used to denote secure federated retrieval and answer selection across private data silos. The system is a secure Federated RAG pipeline built with Flower. Each silo performs retrieval locally over a FAISS index, while server-side aggregation and answer generation run inside a Trusted Execution Environment. The design includes attestation and authorization, TLS-protected communication, and an optional Flower Confidential Remote Compute path for confidential remote LLM inference.

The retrieval pipeline is technically explicit. Each silo embeds the query locally, searches a FAISS `IndexFlatL2` index with metric `faiss.METRIC_L2`, and returns top-$k$ snippets and scores. The per-document dissimilarity is
$$
d(q,d) \triangleq \lVert e(q) - e(d)\rVert_2^2.
$$
The server TEE then performs cross-silo ranking with Reciprocal Rank Fusion, using
$$
\mathrm{rrf\_term}(i) = \frac{1}{k_{\mathrm{RRF}} + i + 1},
$$
with $k_{\mathrm{RRF}}=60$ in the experiments, followed by deduplication and construction of a unified augmented context. Generation can occur inside the server TEE with SmolLM 1.7B Instruct or inside Flower CRC with a large model such as Qwen3 235B.

The system also evaluates a mixed-trust cascade. An external non-confidential model, Amazon Nova Micro in the experiments, is queried for an auxiliary answer, but private documents are not sent to Nova; its output is inserted as auxiliary probabilistic context inside the server TEE. On MIRAGE benchmarks, the reported mean end-to-end times and accuracies are: standalone inference on PubMedQA, $43\mathrm{s}$ and $0.32$; cascaded inference, $48\mathrm{s}$ and $0.45$; confidential inference with CRC, $25\mathrm{s}$ and $0.47$. On MedQA, the corresponding accuracies are $0.39$, $0.57$, and $0.78$, with latencies $43\mathrm{s}$, $46\mathrm{s}$, and $25\mathrm{s}$. In this usage, FedRansel is therefore a systems term for confidential cross-silo retrieval, aggregation, and answer synthesis.

## 5. FedRansel as a naming error for FedeRank

In "FedeRank: User Controlled Feedback with Federated Recommender Systems" [2012.11328], FedRansel is explicitly not a separate algorithm. The paper identifies it as a misspelling or alternative naming of FedeRank. FedeRank itself is a federated learning-to-rank recommender built around a personal factorization model. The server stores item latent factors and item biases, while each client stores its user latent factor and raw interaction vector. The prediction rule is
$$
\hat{x}_{ui}(\Theta) = b_i + p_u^\top q_i,
$$
and training follows a Bayesian Personalized Ranking objective over triples $(u,i,j)$ with $x_{ui}=1$ and $x_{uj}=0$.

The feature that most directly explains the naming confusion is the user-controlled masking of server updates. Each client masks positive-item updates with a randomized mask $\Phi$, governed by a share fraction $\pi \in [0,1]$. Positive-item rows are transmitted with probability $\pi$ and masked with probability $1-\pi$, while negative-item rows are always transmitted. This allows users to control the portion of sensitive signal they share without revealing raw data or local user embeddings.

The reported experiments cover Amazon Digital Music, LibraryThing, and MovieLens 1M. For Amazon Digital Music, FedeRank with $T=X^+/|U|$ reaches $P@10=0.01422$ versus $0.00921$ for centralized BPR-MF; for LibraryThing, the same setting reaches $0.08512$ versus $0.07009$. The paper also reports that accuracy increases rapidly from $\pi=0.0$ to small $\pi$, reaching up to $\approx 92\%$ of the best accuracy even with $\pi=0.1$. In this literature branch, the technically correct referent is FedeRank rather than FedRansel.

## 6. FedRansel as pseudo-centralized random-intersection aggregation

In "A Privacy-Preserving Federated Framework with Hybrid Quantum-Enhanced Learning for Financial Fraud Detection" [2507.22908], FedRansel is a named method. It is the privacy-preserving and robustness layer of a federated framework built around a quantum-enhanced LSTM. The method combines two mechanisms: random parameter sampling at clients and intersection-based robust averaging at the server. Each client trains a local model $w^{t,(k)}$, samples an index set $S_k \subset [1,\dots,|M|]$ of size $\lceil T_l |M|\rceil$, and sends only those coordinates. The server computes the common intersection
$$
C = \bigcap_{k=1}^{K} S_k,
$$
averages only those coordinates,
$$
g_q = \frac{1}{K}\sum_{k=1}^{K} w^{t,(k)}_q \qquad (q \in C),
$$
then samples a return set $G_f$ of size $\lceil T_g |C|\rceil$ and sends it back to clients. The paper calls this regime pseudo-centralized because the server never holds a complete global model.

The surrounding learning problem is binary fraud detection with a global federated objective $\min_w \sum_{k=1}^{K} p_k F_k(w)$ and BCE loss. The recurrent backbone is a QLSTM in which the gate outputs are produced by a variational quantum circuit:
$$
\{f_t,i_t,o_t,g_t\} = \mathrm{VQC}(x_t,h_{t-1}),
$$
with RX angle encoding and Rot+CNOT layers. The experiments use 5 clients, 5 global rounds, 50 local epochs, and $T_l=T_g=0.8$.

The paper emphasizes empirical robustness rather than formal $(\epsilon,\delta)$-DP guarantees. Secure aggregation, TEEs, and homomorphic encryption are not used; privacy derives from randomized subsampling, intersection filtering, and limited server visibility. On Dataset 1, the reported clean-task comparison is: LSTM, Accuracy $0.84$, Recall $0.84$, AUC $0.91$; QLSTM, $0.90$, $0.88$, $0.94$; QLSTM + FedRansel, $0.89$, $0.88$, $0.92$. On Dataset 2, the corresponding values are: LSTM, $0.91$, $0.85$, $0.96$; QLSTM, $0.96$, $0.95$, $0.98$; QLSTM + FedRansel, $0.95$, $0.92$, $0.97$. Under poisoning on QLSTM, Dataset 1 shows baseline degradation of $\sim 6\%$ in Accuracy and $\sim 5\%$ in AUC, while FedRansel reduces these to $\sim 2\%$ and $\sim 1\%$. The paper further reports that FedRansel reduces membership inference attack accuracy by $\approx 4$–$8\%$ relative to the DP baseline.

## 7. Cross-context interpretation

A recurrent misconception is to read FedRansel as a single federated primitive. The literature instead binds the name to different selection operators: client subsets in contribution valuation, LoRA families in VLM fine-tuning, retrieval evidence in federated RAG, masked positive-feedback exposure in recommendation, and intersected parameter coordinates in pseudo-centralized fraud detection [2602.05693] [2503.07216] [2603.25374] [2012.11328] [2507.22908].

These usages also differ in objective. In the contribution-valuation setting, the relevant distinction is between participation randomness and aggregation-rule randomness. In FedRand, the objective is empirical privacy robustness against MIAs while preserving LoRA-based utility and reducing communication. In federated RAG, the focus is confidentiality of prompts, snippets, and generation through TEE-protected aggregation and inference. In FedeRank, the operative mechanism is user-controlled disclosure of positive feedback. In the fraud-detection framework, FedRansel is a robustness-and-privacy layer based on randomized partial visibility and intersection filtering. This suggests that the term does not identify a unique architecture, threat model, or estimator; its meaning is entirely context-dependent.

Source: https://www.emergentmind.com/topics/fedransel