byzantine-robustness

Research testbed for federated learning with hybrid quantum-classical models and post-quantum protection (ML-KEM, ML-DSA, secure aggregation). Reproducible results: a 4-qubit VQC head vs a matched classical head, robust aggregation vs poisoning, PQC cost per round. PennyLane, PyTorch, Flower, FastAPI.

Federated learning + secure aggregation + zero-knowledge norm-bound defence

A benchmark testbed for evaluating backdoor resilience and adversarial robustness in federated skin lesion vision models.

Federated learning sentiment analysis with AT-FedAvg — adaptive trust-aware aggregation across IMDB, Sentiment140 & Amazon. BiLSTM · PyTorch · Flower · FastAPI · Streamlit

Asynchronous Byzantine-Robust Federated Learning with Layer-Stratified Subspace Filtering and Behavioral Memory under Extreme Non-IID Skew.

A bank-agnostic fraud-intelligence layer that learns risk thresholds across banks without any bank sharing its data — with the evaluation that says exactly how far that claim has been tested.

Eight-peer Adult benchmark of non-IID learning, poisoned updates, robust aggregation and membership auditing.

Decentralized federated reinforcement learning with gossip, ADMM consensus and Byzantine robustness

FUSE federated-learning Byzantine screening: faithful reimplementation with a 128-run audit grid across 7 attacks (sign-flip, IPM, Gaussian, ALIE, Min-Max, Min-Sum, magnitude scaling)

Cross-silo federated learning on Fed-ISIC2019 (6 real hospital centres): FedAvg/FedProx/SCAFFOLD/FedAdam, DP-SGD with an independent RDP accountant, membership-inference and poisoning attacks, secure + robust aggregation, 50x update compression, and a real Raspberry Pi 5 edge client over Flower/gRPC.

Siren: Byzantine-robust Federated Learning via Proactive Alarming (SoCC '21)

TrustFed is a federated learning defense system that assigns reputation scores to clients using Cosine Alignment Score (CAS), Loss Change Score (LCS), and Norm Consistency Score (NCS). It detects and down-weights malicious clients in real time, outperforming FedAvg, Multi-Krum, Trimmed Mean, FLTrust, and FoolsGold under four attack types.

Information-theoretic framework for measuring trust in AI systems. Code accompanying Dong (2026), Communications AI & Computing.

Topology-adaptive Byzantine defense for decentralized federated learning — the defense that adapts to your network, not the other way around

Plug-and-play MNIST library for adversarial ML: evasion attacks, backdoor attacks and defenses, and Byzantine-robust federated learning.

Decentralized Continual Cyber Defense: Federated Continual Learning (FCL) for Encrypted Traffic Intrusion Detection with Byzantine Robustness and Differential Privacy

Research prototype for centralized and simulated federated intrusion detection, with hierarchical classification and an explainable pretrained replay demo.

Train one machine-learning model across many parties who don't trust each other. A dishonest minority can't skew the result, and every party computes byte-for-byte identical output on any CPU or architecture (big-endian included) — so each round ships a receipt anyone can re-check offline and reproduce exactly. Rust; drop-in for Flower.

Certified Byzantine robustness for model-heterogeneous federated learning: ensemble subsets + robust distillation aggregation with honest certificates