RIFinder: Signal Extraction in Multiple Domains

• RIFinder is a term for multiple frameworks designed to isolate target signals from confounding structures across wireless, localization, and genomics applications.
• In wireless security, it employs techniques like adversarial training and contrastive pretraining to mitigate receiver and channel variability.
• For localization and genomics, RIFinder utilizes feature selection and statistical validation to manage complex configuration spaces and discordant phylogenies.

RIFinder is not a single canonical method but a recurrent research name applied to distinct technical frameworks in several fields. In the arXiv literature represented here, it denotes at least three substantive usages: a receiver-agnostic radio frequency fingerprint identification framework for LoRaWAN and related IoT settings, a reconfigurable-intelligent-surface-assisted wireless fingerprinting localization framework, and a phylogeny-based pipeline for detecting remote introgression in grasses (Shen et al., 2022, Nguyen et al., 2020, Huang et al., 10 Jul 2025). A later LoRa-focused formulation further uses the name for a three-stage channel-robust and receiver-independent RFFI pipeline based on contrastive pretraining and Siamese fine-tuning (Ma et al., 12 Dec 2025). The shared label therefore refers less to a single lineage than to a family of systems for extracting a target signal from confounding structure.

1. Terminological scope and principal usages

The term has been used for multiple frameworks with different data types, objectives, and inferential targets.

Usage of “RIFinder”	Research area	Representative paper
Receiver-agnostic and collaborative RFFI	Wireless security / IoT authentication	(Shen et al., 2022)
RIS-assisted wireless fingerprinting localization	Indoor positioning	(Nguyen et al., 2020)
Remote introgression detection	Evolutionary genomics	(Huang et al., 10 Jul 2025)
Channel-robust, receiver-independent RFFI	LoRa device authentication	(Ma et al., 12 Dec 2025)

In the wireless literature, RIFinder is associated with disentangling transmitter-specific signatures from nuisance factors such as receiver hardware, channel variation, and SNR heterogeneity. In localization, it refers to a system in which RIS states act as controllable features for RSS fingerprinting. In genomics, it denotes a modular phylogeny-based workflow for detecting gene-tree/species-tree discordance consistent with remote introgression. A common misconception is therefore that “RIFinder” names a single algorithmic artifact. The corpus instead shows a reused name attached to domain-specific frameworks.

2. RIFinder in radio frequency fingerprint identification

In radio frequency fingerprint identification, RIFinder addresses the failure of conventional deep RFFI under receiver drift and receiver change. The 2022 framework models the received signal as

$$y(t) = \mathcal{G}\Big(h(t)*\mathcal{F}^k(x(t))\Big) + n(t),$$

where $\mathcal{F}^k(\cdot)$ is the transmitter chain of DUT $k$, $h(t)$ is the wireless channel, and $\mathcal{G}(\cdot)$ is the receiver hardware effect. The central claim is that receiver hardware alters the learned feature distribution and violates the i.i.d. assumption. To counter this, RIFinder uses adversarial training with a feature extractor, a transmitter classifier, and a receiver classifier connected through a gradient reversal layer. The feature extractor update becomes

$$\theta \leftarrow \theta - \mu \left(\frac{\partial \mathcal{L}_{tx}}{\partial \theta} - \lambda \frac{\partial \mathcal{L}_{rx}}{\partial \theta}\right),$$

so that transmitter discrimination is preserved while receiver-specific information is suppressed (Shen et al., 2022).

That framework also includes collaborative inference. Because the same packet can be captured by multiple receivers, each receiver outputs a probability vector and the outputs are fused by either soft fusion or adaptive soft fusion, the latter weighting each receiver by estimated SNR. In a balanced-SNR setting, using seven receivers improved accuracy by up to 20% over a single-receiver system, and in office experiments with three receivers collaborative inference improved accuracy by over 10% at harder locations. Fine-tuning on a new under-performing receiver with a much smaller learning rate yielded up to 40% improvement, with meaningful gains from as few as 20 packets per DUT-SDR pair and saturation around 50 packets. The LoRaWAN case study used 10 LoRa devices and 20 SDR receivers of six models; heterogeneous adversarial training achieved more than 75% accuracy on all 20 SDRs and more than 95% accuracy on receivers other than RTL-SDRs (Shen et al., 2022).

A later LoRa-focused RIFinder formulation extends this line by combining spectrogram-based representation, unsupervised contrastive pretraining, and Siamese fine-tuning. Its three stages are pretraining, Siamese training, and inference. The model uses spectrograms because, in the log-magnitude STFT domain, transmitter, channel, and receiver effects become approximately additive components. Pretraining uses NT-Xent loss with $\tau=0.05$ and batch size 32; Siamese fine-tuning combines contrastive loss with cross-entropy loss. On three public LoRa datasets and one self-collected dataset, the method reported over 90% accuracy in dynamic non-line-of-sight scenarios when there are only 20 packets per device, and over 85% accuracy across all receivers in LOS and NLOS in the self-collected dynamic multi-receiver setting (Ma et al., 12 Dec 2025).

These RFFI variants are unified by a specific nuisance-removal program: transmitter fingerprints are treated as the invariant object of interest, while receiver and channel effects are modeled as domain shift. The main practical constraints are likewise explicit. Fine-tuning requires support for training/backpropagation at the receiver or gateway, collaborative inference requires multiple receivers and fusion infrastructure, pretraining is computationally heavy, and the strongest receiver-agnostic performance depends on receiver diversity during training (Shen et al., 2022, Ma et al., 12 Dec 2025).

3. RIFinder in RIS-assisted wireless fingerprinting localization

In indoor positioning, RIFinder denotes a framework in which a reconfigurable intelligent surface is used to generate a set of distinguishable RSS radio maps from a single access point. The system contains one transmitter or AP, one RIS, and one mobile user. During localization, the AP sends a burst of $M \le S$ messages, the RIS is reconfigured once per message, and the MU forms an online fingerprint vector

$\mathbf{R}_x = [R_1, R_2, \ldots, R_M].$

The offline database stores, for each sampled location $l$, an RSS vector

$\mathcal{F}^k(\cdot)$0

Localization is then performed by matching the online vector to the database, including a representative rule based on the permuted Pearson correlation coefficient (Nguyen et al., 2020).

The core physical idea is that each RIS configuration produces a distinct propagation pattern, so each RIS state acts as an additional feature in the fingerprint vector. The challenge is combinatorial: with $\mathcal{F}^k(\cdot)$1 quasi-passive tunable elements and $\mathcal{F}^k(\cdot)$2 discrete impedance values per element, the full state space is

$\mathcal{F}^k(\cdot)$3

RIFinder therefore reduces the problem to selecting a small subset of informative RIS states. A heuristic state selection criterion maximizes inter-state RSS dissimilarity over locations, but the paper explicitly notes that this heuristic suffers from outlier bias. The main proposed method instead treats RIS configurations as features and performs wrapper-based feature selection using a genetic algorithm over a surrogate set $\mathcal{F}^k(\cdot)$4, with the fitness objective defined by downstream localization error (Nguyen et al., 2020).

The channel model is not a Friis-style simplification. It is an electromagnetically consistent end-to-end model,

$\mathcal{F}^k(\cdot)$5

with the RIS-mediated term approximated in the far field by a double sum over mutual-impedance terms and RIS coupling coefficients. The resulting received power and RSSI define the radio map. In the reported simulations, $\mathcal{F}^k(\cdot)$6-NN performed best among weighted $\mathcal{F}^k(\cdot)$7-NN, a one-hidden-layer neural network, and random forest. Feature selection consistently improved accuracy, ML-based feature selection outperformed heuristic state selection, and heuristic state selection was the worst method in some settings, even worse than random selection. A mean localization error of about 2 m could be achieved either by $\mathcal{F}^k(\cdot)$8 and $\mathcal{F}^k(\cdot)$9 random RIS configurations or by $k$0 and ML-FS with $k$1, implying a trade-off between map density and RIS-state count (Nguyen et al., 2020).

The significance of this RIFinder usage lies in turning the RIS from a communication-enhancement device into a controllable radio-map generator. Its limits are also clear in the formulation: acquisition cost scales with both the number of reference locations $k$2 and the number of RIS states $k$3, the full RIS state space is intractable, and performance depends heavily on feature selection quality (Nguyen et al., 2020).

4. RIFinder in remote introgression detection

In evolutionary genomics, RIFinder is a modular Python-based workflow for detecting remote introgression, defined as DNA exchange between phylogenetically distant species that falls between standard genomic introgression and cross-kingdom HGT. The workflow has three main stages: reconstruct homologous-gene phylogenies, detect RI-like topological discordance, and apply statistical tests to reject incomplete lineage sorting and other artifacts. Upstream processing uses DIAMOND with E-value $k$4-$k$5 and minimum query coverage of 50%, MCL with inflation parameter 2, MAFFT, TrimAl, ModelFinder, and IQ-TREE with 3,000 ultra-fast bootstrap replicates (Huang et al., 10 Jul 2025).

A key preprocessing stage is gene-tree compression and decomposition. Long-branch outliers are removed using a 99th-percentile threshold on leaf-to-ancestor distances, monophyletic groups from a single defined clade are collapsed, and a tree-splitting algorithm partitions multi-copy homologous-group trees into ortholog group-like subtrees. Candidate RI events are then scored from the arrangement of sister branches, the number of discordant taxa, and the distribution of those taxa in the tree. The threshold is stated to be determined using Cayley’s formula, $k$6, where $k$7 is the number of taxa in a sub-clade. Only nodes with bootstrap over 70 are retained by default, and alternative rootings can be rescored to reduce rooting bias (Huang et al., 10 Jul 2025).

To distinguish RI from ILS, RIFinder applies a modified Branch-Length Test. For each acceptor, donor, sister triplet, it compares the distributions of $k$8 and $k$9 with a two-tailed independent t-test. Under the null, these distances are comparable; under RI, the donor-acceptor pair is more recent. Genome-wide support is accepted after Benjamini–Hochberg correction at FDR $h(t)$0. For locus-specific validation, the framework uses RELL, SH, and AU tests in IQ-TREE with 10,000 bootstrap replicates, rejecting alternative topologies at $h(t)$1 (Huang et al., 10 Jul 2025).

On simulations, RIFinder achieved precision of 85.7% to 98.6%, and at a 0.1% transfer rate recall was 89.0%. Applied to 122 grass genomes, it detected 622 candidate RI events from 543 distinct homologous genes, including 286 singleton, 64 doubleton, and 272 multiple-species events. The broad pattern was asymmetrical, with transfers more frequent from PACMAD to BOP than in the reverse direction. The paper highlights a Triticeae-specific RI pulse around 30 mya, a $h(t)$2-kbp RI segment in Cleistogenes songorica, and RI signals involving gramine biosynthetic gene clusters. The authors also state that distinguishing RI from ILS remains difficult, that the current framework has limited phylogenetic resolution, and that the 622 events are likely a lower-bound estimate (Huang et al., 10 Jul 2025).

5. Cross-domain methodological structure

Despite their disciplinary separation, the main RIFinder usages share a comparable systems pattern. Each builds a representation in which a latent target variable is separated from nuisance structure, then uses that representation for a downstream decision. In RFFI, the invariant target is the transmitter fingerprint, while receiver and channel effects are suppressed through adversarial learning, contrastive pretraining, or Siamese pairing. In RIS-assisted localization, the target is spatial identity, and RIS states are treated as controllable features whose subset must be optimized to maximize location discriminability. In RI detection, the target is deep inter-clade gene flow, and large homologous families are compressed and split so that discordant topologies can be isolated and filtered statistically (Shen et al., 2022, Nguyen et al., 2020, Huang et al., 10 Jul 2025, Ma et al., 12 Dec 2025).

This suggests that “RIFinder” functions less as a narrow algorithmic label than as a naming convention for structured inference under confounding. That inference should remain qualified: the papers do not claim a shared genealogy across fields. What they do show is a recurring architecture of preprocessing, candidate formation, nuisance control, and validation.

6. Misconceptions, limitations, and significance

The first misconception is nominal. RIFinder is not a single universally recognized framework. It is a reused title applied to unrelated technical systems in wireless security, indoor localization, and comparative genomics. Treating results from one usage as transferable to another would therefore be category error.

The second misconception is methodological. In wireless RFFI, receiver hardware is not merely a minor implementation detail. The 2022 LoRaWAN study reports that conventional training on RTL-SDR and testing on different receivers could drop to about 20% accuracy on some receivers, and that low-cost RTL-SDRs can lose more than 40% accuracy over four days. Receiver variation is therefore a primary domain-shift mechanism, not an afterthought (Shen et al., 2022). The later LoRa RIFinder further emphasizes that spectrogram representation, contrastive pretraining, and Siamese fine-tuning are required to jointly mitigate channel and receiver impairments when labeled data are scarce (Ma et al., 12 Dec 2025).

In RIS localization, the main limitation is not merely classifier choice but state-space management and radio-map acquisition burden. The full configuration space $h(t)$3 is enormous, heuristic state selection can be biased by outliers, and the cost of database construction scales with both $h(t)$4 and $h(t)$5 (Nguyen et al., 2020). In RI detection, the central difficulty is causal attribution: discordant gene trees may arise from ILS, duplication, loss, or sampling artifacts, so the method deliberately prioritizes specificity over sensitivity, yielding a conservative set of events (Huang et al., 10 Jul 2025).

Taken together, these usages establish RIFinder as a polysemous but technically coherent research label. Across domains, it denotes frameworks designed to recover informative structure from observations in which the signal of interest is entangled with nuisance variation, combinatorial complexity, or conflicting histories.