Source-linked AI summary
Deep Learning-Based Multi-User Communication Design for Dense IoT Networks: Interference-Aware Finite-Blocklength Communication and Preliminary MIMO Extensions
Arkadeep Sinha, Shubham Paul, R. Manivasakan
TL;DR
Dense IoT communication must handle finite-blocklength packets under multi-user interference while keeping receiver complexity manageable. The paper extends SiameseNet to multi-user end-to-end learned transceivers, evaluates robustness and latent geometry, and reports preliminary 2×2 MIMO results. The framework retains strong BLER performance and scalable per-user decoding, while MIMO conclusions remain limited by fixed-channel evaluation.
Problem
Dense IoT networks need reliable finite-blocklength communication under spectrum, latency, reliability, and multi-user interference constraints, while scalable low-complexity detection remains challenging.
Method
The paper extends the two-user SiameseNet transceiver to 4 and 8 users, evaluates interference robustness, analyses latent codeword geometry, and explores preliminary 2×2 MIMO settings.
Results
The framework scales gracefully, retains strong BLER performance under interference variations, outperforms NOMA and RSMA under unequal interference strengths, and shows increasing minimum distance as effective per-user rate decreases.
Takeaways & Limitations
Learned multi-user transceivers offer a scalable approach for interference-limited IoT communication, with latent-space separation providing geometric insight into observed BLER gains.
Abstract
from arXiv · showhide
Dense IoT networks require reliable communication despite limited spectrum and substantial multi-user interference while maintaining manageable receiver complexity. This work introduces a deep-learning-based end-to-end multi-user communication design for interference-limited finite-blocklength IoT scenarios, focusing on short and medium blocklengths. We extend a prior 2-user SiameseNet transceiver framework to accommodate 2, 4, and 8 users, leveraging learned redundancy for interference suppression and noise robustness. Compared to conventional non-orthogonal access baselines, our method demonstrates strong Block Error Rate (BLER) performance across various scenarios without resorting to joint detection; the per-user decoder scales roughly linearly with the number of users. Further, we examine the robustness under interference mismatch and unequal interference strengths, critical for practical deployments with heterogeneous devices. The Latent-space analysis reveals that the learned codeword distance increases as the effective per-user rate decreases, corroborating with the observed BLER improvements. In addition, we also present preliminary results for a 2X2 MIMO setup under fixed-channel CSIT and CSIR, indicating potential for extending the framework to IoT gateways with multiple antennas.
I. INTRODUCTION
Dense IoT networks must support short finite-blocklength packets under spectrum, latency, reliability, noise, and multi-user interference constraints. This paper extends learned multi-user communication beyond two users, evaluates robustness and complexity trade-offs, and studies latent geometry and preliminary MIMO extensions.
- Dense IoT packets face substantial multi-user interference alongside stringent spectrum, latency, reliability, and finite-blocklength constraints.
- Orthogonal access and FEC can reduce spectral efficiency, while NOMA and RSMA may introduce complex joint detection that challenges scalable IoT receivers.
- The paper generalises SiameseNet from two users to 4-user and 8-user interference-limited settings and systematically evaluates BLER.
- The study tests robustness to mismatched and unequal interference strengths, reflecting heterogeneous multi-device deployments.
- Latent codeword geometry is analysed through minimum-distance and correlation measures to explain observed BLER trends.
- Preliminary 2×2 MIMO results extend the framework beyond SISO under fixed-channel assumptions.
II. SYSTEM MODEL
The system extends a two-user SiameseNet interference model to N users, where each user encodes a message into a codeword and independently decodes its received vector.
- The interference model supports N ∈ {4, 8} users, with symmetric interference magnitudes satisfying αij = αji.
- Each user maps a k-bit message Mi to a real n-length codeword zi through a parameterised encoder Ei.
- The receiver obtains a vector yi containing the transmitted codeword, interference from other users, and white noise.
- Each decoder Di independently estimates user i’s message from yi using decoder parameters ϕi.
III. SIAMESE ARCHITECTURE AND DECODING COMPLEXITY
The Siamese architecture trains encoders jointly through shared interference-related losses while learning each user’s decoder independently, with interference injected before decoding.
- Model Structure and Design: The autoencoder combines an encoder, a non-trainable channel layer modelling latent perturbations, and a decoder.
- Model Structure and Design: The encoder uses dense, linear, batch-normalisation, and Gaussian layers, while the decoder ends with a softmax output after interference is added at the channel input.
- Model Structure and Design: For N users and blocklength L, hidden and encoder-output sizes are NL, and each decoder has a hidden layer of size NL.
- Training: Extending the Siamese network to N users allows a change in encoder parameters θi to affect all user costs, enabling joint encoder learning.
- Training: Decoder parameters ϕi are learned independently because perturbations to ϕi affect only user i’s cost.
B. Decoder Complexity
The proposed per-user decoder operates on one received vector with complexity linear in the user count, avoiding the exponential search required by joint maximum-likelihood detection.
- Decoder Complexity: O(Nn) is the decoder forward-pass complexity, because user i processes yi with a hidden layer proportional to Nn.
- Training: The training algorithm iteratively encodes messages, forms received vectors with interference, updates decoders from individual costs, and updates encoders from all-user costs.
- Decoder Complexity: The complexity comparison is summarised as per-receiver decoding complexity for N ∈ {2, 4, 8}.
- Decoder Complexity: Joint maximum-likelihood detection requires O(2^(kN)) evaluations, while NOMA-SIC uses N successive stages with inter-stage error propagation.
IV. BLER PERFORMANCE
The study evaluates finite-blocklength BLER for 2, 4, and 8 users against TDMA, NOMA, and RSMA baselines under matched effective rates and interference settings. It also considers SIC-based decoding and practical baseline limitations, including unequal-interference suitability and the indicative role of the PPV benchmark.
- Experimental setup: Experiments cover (k, n) settings of (4, 8), (16, 64), and (256, 2048) for N = 2, 4, and 8, spanning short and medium blocklengths.The evaluation varies Eb/N0 from 0 to 8 dB.
- Baselines: TDMA uses uncoded BPSK with rate 1/N, while baselines are configured so both systems have an overall rate of 1 bps/Hz.The SiameseNet design and baselines transmit k information bits per user with matched effective rate R = 1/N.
- Baseline limitations: Under symmetric powers, NOMA uses a random interference-cancellation order, which may underestimate performance relative to power-asymmetric settings with optimized ordering.The paper notes that NOMA is more suited to unequal interference strengths and treats it as an indicative baseline for Experiment I.
- Baselines: NOMA and RSMA provide SIC-based references: NOMA cancels interfering streams before decoding the desired stream, whereas RSMA removes common streams before decoding the private stream amid residual interference.RSMA splits each message into common and private parts and uses a decoding order for common streams.
- Finite-blocklength benchmark: The PPV normal approximation is used only as an indicative single-user AWGN finite-blocklength benchmark because a detailed information-theoretic analysis is beyond the paper’s scope.Its purpose is to contextualize the performance gap introduced by multi-user interference.
B. Experiment I: Equal Interference from All Interferers
Experiment I studies equal interference for 2, 4, and 8 users. SiameseNet outperforms TDMA, with larger gains as the per-user rate decreases, while comparison with PPV shows stronger suitability for short blocklengths than larger ones.
- BLER versus SNR: 0.5 dB gain is observed for 2 users versus TDMA, while the 4-user gain reaches about 2.5 dB and the 8-user case improves further.All systems have the same overall rate of 1 bps/Hz, with per-user rates decreasing from 1/2 to 1/8 as N increases.
- BLER versus SNR: Lowering the per-user rate from 1/2 to 1/8 increases the implicit minimum distance and improves BLER beyond a crossover SNR.The comparison covers N ∈ {2, 4, 8} under equal interference.
- PPV comparison: Against PPV, SiameseNet is around 0.5 dB worse for 2 users and 1.5 dB worse for 4 users, and much worse for 8 users.The PPV comparisons use rates 1/2, 1/4, and 1/8 for the corresponding user counts.
- PPV comparison: The PPV comparison indicates that the system is more suited to short-blocklength systems than to larger-blocklength systems.
C. Experiment II: Mismatched α: A Test for Robustness
The experiment tests a model trained at α=1 under mismatched interference strengths. Performance remains reasonably robust, with mismatch effects depending on the tested α and SNR.
- The model trained at α=1 is evaluated under different interference-strength values.The experiment compares mismatched testing with correctly matched models.
- The mismatched α=0.1 model performs worse than the correctly trained α=0.1 model, but the gap narrows as SNR increases.
- At α=10, the mismatched model performs as well as the correctly trained model.
- The results indicate that the model can accommodate moderate interference-estimation errors.
D. Experiment III: Unequal Interference Strengths
The unequal-strength experiment evaluates users exposed to different interference levels. Under this setting, the proposed model outperforms NOMA and RSMA, while the broader discussion remains future work.
- For α12=α21=1 and the remaining listed interference coefficients equal to 0.1, user 2 interferes more strongly with user 1 than users 3 and 4 do.
- The experiment compares high- and low-interference users across the proposed model, NOMA, and RSMA.
- The proposed model outperforms both NOMA and RSMA in the unequal-interference scenario.
- A more thorough discussion of unequal interference strengths is left for future work.
V. PRELIMINARY RESULTS WITH MIMO
The paper presents a proof-of-concept 2×2 MIMO extension under fixed Rayleigh channels, considering CSIT and CSIR variants. The reported results show favorable fixed-channel performance, but do not establish general behavior over random channel ensembles.
- MIMO setup: The MIMO study uses a 2×2 interference channel with i.i.d. complex Gaussian links, α=1, and channels held fixed during training.
- CSIT: With CSIT, the encoder receives channel information to support beamforming and coded spatial multiplexing.
- CSIT: 1 bpcu per user is achieved for two simultaneous users, with diversity behavior consistent with fourth order under the fixed-channel setup.
- CSIR: With CSIR, the decoder receives channel information, while the model uses channel-agnostic codewords and suppresses cross-user interference through increased inter-user separation.
- Scope: These MIMO results are preliminary because the channels were drawn once and held fixed rather than evaluated over randomly varying ensembles.
VI. LATENT SPACE ANALYSIS
The latent-space analysis links BLER behavior to learned codeword geometry. Minimum distance increases as effective per-user rate decreases, while orthogonalization is selective and correlation-reduction behavior is not explicitly guaranteed.
- Minimum distance dmin is the most informative geometric indicator of BLER in the reported experiments.
- dmin increases as effective per-user rate decreases, with a stronger effect at α=0.1.
- Average codeword angles cluster near 90°, with same-user and cross-user angle patterns depending on α.
- The model reduces cross-user correlation as interference scales, consistent with the BLER trends.
- The model was not explicitly designed to reduce correlation, and the paper claims no guarantee that it will do so.