Adopting artificial intelligence (AI) in radio access networks (RANs) presents several challenges, limited availability of link-level measurements (e.g., CQI reports), stringent real-time processing constraints (e.g., sub-1ms per TTI), and network heterogeneity (different spectrum bands, cell types, and vendor equipment). A critical yet often overlooked barrier lies in the computational and memory limitations of RAN baseband hardware—particularly in legacy 4thGeneration (4G) systems—which typically lack on-chip neural accelerators. As a result, only lightweight AI models (under 1Mb and sub-100µs inference time) can be effectively deployed, limiting both their performance and applicability. However, achieving strong generalization across diverse network conditions often requires large-scale models with substantial resource demands. To address this trade-off, this paper investigates policy distillation in the context of a reinforcement learning–based link adaptation task. We explore two strategies: single-policy distillation, where a scenario-agnostic teacher model is compressed into one generalized student model; and multi-policy distillation, where multiple scenario-specific teachers are consolidated into a single generalist student. Experimental evaluations in a high-fidelity, 5thGeneration (5G)-compliant simulator demonstrate that both strategies produce compact student models that preserve the teachers’ generalization capabilities while complying with the computational and memory limitations of existing RAN hardware.

In this work, we present polar code designs that offer a provably optimal solution for biometric identification and authentication systems under noisy enrollment for certain sources and observation channels. We consider a discrete memoryless biometric source and discrete symmetric memoryless observation channels. It is shown that the proposed polar code designs achieve the fundamental limits with privacy and secrecy constraints. Depending on how the secret keys are extracted and whether the privacy leakage rate should be close to zero, we consider four related setups, which are (i) the generated secret key system, (ii) the chosen secret key system, (iii) the generated secret key system with zero leakage, and (iv) the chosen secret key system with zero leakage. For the first two setups, (i) and (ii), the privacy level is characterized by the privacy leakage rate. For the last two setups (iii) and (iv), private keys are additionally employed to achieve close to zero privacy leakage rate. In setups (i) and (iii), it is assumed that the secret keys are generated, i.e., extracted from biometric information. While in setups (ii) and (iv), secret keys provided to the system are chosen uniformly at random from some trustful source. This work provides the first examples of fundamental limits-achieving code designs for identification and authentication. Moreover, since the code designs are based on polar codes and many existing works study low-complexity and short block-length polar coding, the proposed code designs in this work provide the code design structure and a framework for the application of biometric identification and authentication.

In this work, we present a polar code design that offers a provably optimal solution for biometric identification systems allowing authentication under noisy enrollment with secrecy and privacy constraints. Binary symmetric memoryless source and channels are considered. It is shown that the proposed polar code design achieves the fundamental limits and satisfies more stringent secrecy constraints than previously in the literature. The proposed polar code design provides the first example of a code design that achieves the fundamental limits involving both identification and authentication.

In this paper, we study fundamental trade-offs in privacy-preserving biometric identification systems with noisy enrollment. The proposed identification systems include helper data, secret keys, and private keys. Helper data are stored in a public database and used for identification. Secret keys are either stored in a secure database or provided to the user, and can be used in a next step, e.g. for authentication. Private keys are provided by users, and are also used for identification. In this paper, we impose a noisy enrollment channel and an arbitrarily small privacy and secrecy leakage rate. We characterize the optimal trade-off among the identification, secret key, private key, and helper data rates. Depending on how secret keys are produced, we study two cases of the proposed privacy-preserving identification systems, where the secret keys are generated and chosen respectively. By introducing private keys, it is shown that the identification system achieves close to zero privacy leakage rate in both generated and chosen secret key settings. The results also show that the identification rate and the secret key rate can be enlarged by increasing the private key rate. This work provides a framework for analyzing privacy-preserving identification systems and an insight on the design of optimal systems.