Quantum-Safe Federated Learning: PQC-Protected Secure Aggregation with ML-KEM and ML-DSA

Samuel Yen-Chi Chen, Shinjae Yoo

Sujith B

Federated Learning (FL) enables collaborative machine learning without centralizing raw data. However, the impending threat of cryptographically-relevant quantum computers exposes FL's reliance on classical public-key cryptography (e.g., RSA, ECC) to Harvest-Now-Decrypt-Later (HNDL) attacks. We present PQ-FL, the first open federated learning control plane secured end-to-end using the finalized NIST Post-Quantum Cryptography (PQC) standards. Our architecture integrates ML-DSA-87 (FIPS 204) for transport-layer identity authentication (mTLS) and ML-KEM-1024 (FIPS 203) for application-layer model payload encryption. To overcome the lack of comprehensive PQC support in standard Python libraries (BoringSSL), we introduce a novel OQS-stunnel sidecar pattern that proxies edge client traffic. We evaluate the system on a CUDA-accelerated edge cluster training a convolutional neural network. Empirical results confirm successful multi-round aggregation using FedAvg, Krum, and Trimmed-Mean strategies alongside differential privacy, achieving seamless PQC key encapsulation (1568-byte ciphertexts) and verifiable asynchronous global model streaming.

Aniello Castiglione, Vincenzo Loia, Michele Nappi +2 more

Industrial innovations, driven by technologies such as the Internet of Things (IoT), cloud computing, and Artificial Intelligence (AI), have revolutionized various operational domains, including supply chain management, equipment monitoring, predictive maintenance, and quality control. Deep Learning (DL) has become instrumental in addressing complex, data-driven challenges within industrial settings, relying on large datasets collected from heterogeneous sources. However, centralizing these datasets on cloud platforms poses significant privacy and security risks. Federated Learning (FL) offers a promising solution by enabling distributed model training across multiple nodes while preventing the exchange of raw data. Despite its advantages, FL introduces new vulnerabilities, especially related to the security of communication channels between participating entities. Traditional cryptographic mechanisms, such as digital signatures and model encryption, can mitigate these risks, but the emergence of quantum computing threatens the robustness of conventional solutions. This work explores the integration of Post-Quantum Cryptography (PQC) into FL to enhance security without incurring significant performance degradation. A modular FL architecture with three security levels is proposed: (i) an unprotected baseline; (ii) authenticated communication using ML-DSA digital signatures; and (iii) full protection combining ML-DSA and ML-KEM-512-based key encapsulation. Implemented within the Flower framework, the architecture is evaluated under simulated MitM attacks. Experimental results demonstrate that PQC-enhanced schemes effectively mitigate quantum-resistant threats while maintaining acceptable computational overhead, thereby ensuring model integrity and data confidentiality.

S Ravikumar, E. Chandralekha, K. Vijay +2 more

Dev Gurung, Shiva Raj Pokhrel, Gang Li

Vikram Choudhary

The emergence of cryptographically relevant quantum computers (CRQCs) would compromise today’s dominant public-key primitives (RSA, DiffieHellman, ECDSA), threatening long-lived confidentiality, authentication, and integrity guarantees across AI pipelines and blockchain networks. This manuscript proposes a unified, quantum-resilient architecture that blends NISTstandardized post-quantum cryptography (PQC)—MLKEM for key establishment and ML-DSA or SLH-DSA for digital signatures—with defense-in-depth controls for AI model protection, data privacy, and blockchain protocol evolution. We first motivate the need for migration by outlining quantum attacks (Shor for factoring and discrete logarithms; Grover for quadratic search speedups) and summarize PQC standardization milestones and transition guidance. We then review the literature on lattice-based and hash-based cryptography, homomorphic encryption for privacy-preserving AI, and secure aggregation/differential privacy as complementary privacy technologies. Building on this foundation, we present a methodology for designing and evaluating quantum-resistant AI and blockchain systems, including key-lifecycle hardening, protocol-level hybrids, and crypto-agility. A simulated deployment (1,000 clients; 16 validator nodes; TLS 1.3 with ML-KEM and hybrid handshakes; blockchain transactions signed with ML-DSA and SLH-DSA) measures handshake and signing overheads, throughput, and model-training privacy-utility trade-offs. Results indicate that: (i) PQC-only handshakes add modest latency yet remain acceptable for online APIs; (ii) hybrid ECDHE+ML-KEM offers near-term defense against “store-now-decrypt-later,” with predictable cost; (iii) ML-DSA is practical for blockchain validation at common block sizes, while SLH-DSA suits firmware/codesigning and archival integrity; and (iv) combining PQC with secure aggregation and differentially private training preserves model utility with bounded overhead.