SSL-FL: A lightweight authentication framework for secure federated learning

Federated learning (FL) enables distributed model training with data privacy, but requires efficient authentication and communication in resource-constrained environments. This paper presents SSL-FL, a framework that combines semiconductor superlattice physically unclonable functions (SSL-PUFs) and compressed sensing (CS). SSL-PUFs exploit chaotic synchronization to generate device-specific true random keys from quantum randomness, enabling decentralized authentication without centralized key management or large storage tables. CS compresses model updates, reducing transmission overhead while retaining data required for aggregation. The protocol employs multi-client challenge perturbations to enforce cross-device key generation consistency against side-channel attacks. An adaptive timing mechanism further optimizes synchronization under variable network conditions. Security is verified with BAN logic and Real-Or-Random models, confirming the consistency and indistinguishability of keys, and addressing various threats through informal analysis. Experiments show that it can resist a wide range of attacks from simple replay to complex time manipulation while maintaining moderate accuracy at compression ratios from 0.04 to 0.5. The framework generates proofs of size $$2|C|+|R|+|T|$$ (132 bytes) with computational cost of $$2T_{\text {PUF}}+T_{SE}$$ per round, reducing communication and computation requirements compared to existing approaches. SSL-FL provides a secure and scalable solution for FL in dynamic, resource-limited environments.