Source-linked AI summary
Rethinking the Foundations of Two-Sided AI Models for 6G
Yongjeong Oh, Zihan Chen, Timothy J. O'Shea, Junyong Shin, Jinho Choi, Yo-Seb Jeon, Jihong Park
TL;DR
Two-sided AI models face deployment gaps in legacy coexistence, channel adaptation, and multi-vendor updates. The paper addresses these with NR integration, adaptive model selection, and gradient-free fine-tuning, achieving coexistence and strong adaptation while reducing communication overhead.
Problem
Two-sided AI models lack practical solutions for legacy coexistence, diverse channel adaptation, and multi-vendor updates without exposing proprietary information.
Method
The paper integrates two-sided processing into NR, builds an adaptive model-selection table using surrogate channels, and applies gradient-free zeroth-order fine-tuning.
Results
More than 15,000× lower Vendor B communication overhead is achieved with comparable CSI reconstruction NMSE to conventional Type 2 fine-tuning.
Takeaways & Limitations
These alternatives support moving two-sided AI models toward interoperable and deployable components of next-generation air interfaces.
Takeaways & Limitations
Multi-vendor fine-tuning remains challenging because existing approaches may expose proprietary model information and enable inference or adversarial attacks.
Abstract
from arXiv · showhide
For next-generation air interfaces, two-sided artificial intelligence (AI) models have received growing attention, with AI models deployed at both the transmitter and receiver for efficient channel feedback and data communication. However, their practical deployment is complicated by assumptions commonly made in existing studies, including isolation from legacy users, training under predefined channel conditions, and gradient-based fine-tuning requiring substantial cross-vendor communication. This article revisits these assumptions and presents practical alternatives. First, for legacy coexistence, we integrate two-sided model processing into the 5G New Radio (NR) protocol stack and validate its operation alongside conventional NR on a real-world testbed. Second, instead of training under a massive number of predefined channel conditions, we construct a compact model table by jointly optimizing two-sided models with trainable surrogate channels, and select the best model according to the current channel condition to enable channel adaptation with high task performance and low training/storage overhead. Finally, unlike existing fine-tuning that exchanges large gradient vectors containing potentially private model information, we present gradient-free zeroth-order fine-tuning that requires only scalar feedback, facilitating multi-vendor interoperability. Together, these approaches advance the practical deployment of two-sided AI models while highlighting key open challenges.
I. INTRODUCTION
Two-sided AI models are increasingly integrated into transmitters and receivers, but their practical deployment inherits assumptions about legacy-user isolation, predefined channel training, and backpropagation-based fine-tuning. The article therefore rethinks these foundations through practical alternatives for coexistence, channel adaptation, and model updates.
- Background: Two-sided AI architectures place AI models at both the transmitter and receiver and support directions including neural transceivers, joint source-channel coding, and semantic communication.These research directions differ in objectives and system settings but share a transmitter–receiver model architecture.
- Prevailing assumptions: Existing two-sided-model foundations assume isolation from legacy UEs, training under predefined channel conditions, and backpropagation-based fine-tuning.These assumptions were established in controlled academic studies and carried into practical deployment efforts such as CSI feedback.
- Practical challenges: Legacy isolation prevents coexistence with conventional radio access technologies, while prescribed channel training creates specialization–generalization trade-offs and increases costs when channel-specific models are maintained.A single model trained across diverse conditions can instead compromise channel adaptation, according to the introduction’s problem statement.
- Article scope: The article revisits these assumptions by examining practical alternatives for coexistence with 5G NR, adaptation across diverse channels without channel-specific training, and transmitter–receiver model updates.These three themes define the article’s stated focus for rethinking two-sided AI model deployment.
II. FROM FOUNDATIONAL ASSUMPTIONS TO PRACTICAL LIMITATIONS
The section identifies practical limitations in two-sided AI model deployment involving legacy coexistence, channel adaptation, and multi-vendor interoperability, motivating practical alternatives.
- Legacy Coexistence: AI-native 6G operation must coexist with legacy 5G NR, requiring validation of simultaneous two-sided and conventional NR operation with full-stack compatibility.The validation should span PHY processing through higher-layer procedures while preserving performance advantages.
- Channel Adaptation: Existing channel adaptation commonly trains separate models for predefined channel conditions, creating a one-model-fits-one approach that selects among prepared models.Channel adaptation is identified as a key aspect of AI model life cycle management.
- Multi-Vendor Interoperability: Cross-vendor deployment is difficult because gradient-based joint training requires exchanging models, intermediate features, or gradients between transmitter and receiver sides.3GPP classifies these approaches as Type 1 and Type 2, with different locations for training and model exchange.
- Motivation: These deployment challenges motivate practical alternatives for legacy coexistence, channel adaptation, and multi-vendor interoperability.The alternatives are presented in the following sections.
III. INTEGRATION INTO THE NR PROTOCOL STACK FOR LEGACY COEXISTENCE
The section presents a capability-dependent PHY integration architecture for introducing two-sided model processing alongside conventional NR. A full-stack real-world testbed validates coexistence, protocol operation, and uplink evaluation while preserving legacy NR processing.
- Architecture: A capability-dependent PHY architecture adds two-sided model processing as an alternative to selected conventional NR PHY functions while retaining the unmodified conventional processing path.This design aims to preserve the two-sided model’s performance advantage while supporting legacy coexistence.
- Testbed implementation: The full-stack testbed uses DeepSig OmniPHY Axon, an OCUDU-based gNB on an NVIDIA DGX Spark, a software-defined UE, and a USRP B210 radio front-end.The radio operates in NR band n78 with 20 MHz bandwidth, 30 kHz subcarrier spacing, and split 8.
- Capability-dependent processing: The gNB selects conventional or two-sided processing according to UE capability, serving legacy UEs through the existing NR chain and replacing QAM mapping and pilot insertion for model-enabled UEs.The two-sided UE model maps coded bits directly to complex-valued symbols while retaining NR resource-element mapping.
- Full-stack operation: The integrated system supports UE registration, protocol data unit session establishment, and real-time Internet Protocol traffic, including ping, secure shell, and throughput measurements.Measured roundtrip times are 10–20 ms, and the experiments demonstrate full-stack operation while preserving conventional NR processing for legacy UEs.
- Performance evaluation: Uplink evaluation compares conventional NR and the two-sided model across SNRs using a 3GPP TR 38.901 rural macro channel with a 1 × 4 antenna configuration.The conventional NR baseline uses pilot-aided LMMSE-Wiener channel estimation and an LMMSE equalizer.
IV. ADAPTIVE MODEL SELECTION FOR CHANNEL ADAPTATION
The adaptive model selection (AMS) framework enables channel adaptation by selecting among a compact offline-constructed table of models trained with surrogate channels. On CIFAR-10 image reconstruction, AMS achieves the highest PSNR across 0–30 dB SNR while avoiding the overhead and degradation tradeoffs of baseline schemes.
- Adaptive model selection: AMS selects the model and transmission parameters best matched to the current channel from a compact offline-constructed table.It replaces separate pre-training for every predefined channel condition with surrogate channels (SCs).
- Surrogate-channel optimization: Trainable SCs and two-sided models are jointly optimized using task loss plus SC regularization to avoid convergence to an error-free channel.Analog SCs use trainable noise variances, while digital SCs use trainable bit-flip probabilities.
- Online channel adaptation: During online operation, AMS reproduces each trained SC over the wireless channel by matching target SNR or bit error rate through transmission-parameter control.Feasible entries are identified under constraints such as total transmit power, and the highest-performing feasible entry is selected.
- Performance evaluation: AMS achieves the highest PSNR across the entire 0–30 dB SNR range for CIFAR-10 image reconstruction under Rayleigh fading with 16-QAM.The evaluation uses 64 subcarriers and compares one-model-fits-one, one-model-fits-all, and naïve AMS baselines.
- Performance evaluation: Naïve AMS has overhead comparable to AMS, but its PSNR is 6–8 dB lower over the 0–20 dB range.One-model-fits-one has the second-highest PSNR but substantial training and storage overhead, while one-model-fits-all has the lowest overhead and lowest PSNR over most of the range.
V. GRADIENT-FREE FINE-TUNING FOR MULTI-VENDOR INTEROPERABILITY
The proposed gradient-free zeroth-order fine-tuning enables multi-vendor joint adaptation using forward propagations and scalar loss feedback instead of exchanging parameters or gradients. It limits proprietary-information leakage while achieving comparable CSI reconstruction NMSE with substantially lower memory and communication overhead.
- Framework: Gradient-free fine-tuning replaces parameter or gradient exchange with forward propagations and scalar loss differences between transmitter- and receiver-side vendors.This addresses privacy, memory, and communication concerns associated with 3GPP Type 1 and Type 2 fine-tuning.
- Framework: The framework supports multiple model perturbations in one batch, avoiding additional communication rounds.Perturbed outputs are transmitted together for parallel evaluation.
- Privacy and efficiency: The proposed method prevents this information leakage by eliminating gradient feedback while retaining CSI reconstruction NMSE comparable to conventional Type 2 fine-tuning.The evaluation uses CSI feedback in an urban macro cell scenario with a 32 × 32 angular-delay CSI matrix.
- Privacy and efficiency: Gradient feedback can expose the receiver-side Jacobian and model output: NMSE values are −58.24 dB and −14.56 dB, respectively.These estimates are achieved by Vendor A during Type 2 fine-tuning at a compression ratio of 1/4.
- Privacy and efficiency: 3.65 times lower Vendor A-side memory usage and more than 15,000 times lower Vendor B-side communication overhead are achieved.These reductions accompany CSI reconstruction NMSE comparable to conventional Type 2 fine-tuning.
VI. OPEN CHALLENGES AND FUTURE DIRECTIONS
The paper identifies open challenges for deploying two-sided AI models, including unified AMC-AMS adaptation, extension to task-oriented communication, and multi-vendor lifecycle management. Addressing these challenges requires validating legacy coexistence, managing cross-layer information exchange, and establishing vendor-agreed rules.
- Open Challenges and Future Directions: These challenges extend the paper’s practical alternatives for legacy coexistence, channel adaptation, and multi-vendor interoperability into future research directions.The preceding advances motivate new research directions while highlighting further challenges for practical deployment.
- AMC-AMS Unification: Jointly coordinating AMC and AMS using CQI feedback is an important challenge while guaranteeing legacy coexistence.AMC selects modulation and coding from the MCS table, whereas AMS selects a model-SC pair from the MSC table.
- Extension to Task-Oriented Communication: Extending two-sided models to task-oriented communication is promising but non-trivial because application-layer data and RAN channel information require costly cross-layer exchange.Joint model-SC training avoids this exchange during training, but online operation still requires application-layer visibility.
- Multi-Vendor LCM: Multi-vendor lifecycle management must resolve MSC table sizing, updates, fine-tuning triggers, and vendor responsibilities through standardized or vendor-agreed rules.Transmitter- and receiver-side tables can let vendors maintain separate models while sharing only the SCs.
VII. CONCLUSION
The article addresses deployment challenges of two-sided AI models through NR protocol-stack integration, MSC-table-based AMS, and gradient-free zeroth-order fine-tuning. It identifies open challenges whose resolution is needed for interoperable, deployable next-generation air-interface components.
- Practical deployment alternatives: Three practical alternatives address legacy coexistence, channel adaptation, and multi-vendor interoperability: NR protocol-stack integration, MSC-table-based AMS, and gradient-free zeroth-order fine-tuning.These challenges are examined as key deployment barriers for two-sided AI models.
- Open challenges: Open challenges include unifying AMS and AMC, extending task-oriented communication, and enabling multi-vendor LCM.The article discusses these as remaining challenges after presenting the three practical alternatives.
- Path to deployment: Coordinated advances in research, prototyping, and standardization are critical to moving two-sided models beyond isolated demonstrations toward interoperable and deployable next-generation air-interface components.The conclusion frames coordination across these activities as necessary for practical progress.