Abstract: Byzantine attack poses a critical barrier to deploying blockchain-based Federated Learning (FL) in security-sensitive domains. In a Byzantine attack, a malicious participant can implant a stealthy backdoor while maintaining the model’s apparent utility on clean inputs. Existing Byzantine-resilient blockchain–FL systems often emphasize auditability and participation management. However, these existing blockchain-based FL systems still rely on single-view robust aggregation rules (e.g., distance, norm, or median-style filtering). These single-view aggregation rules are not designed to withstand modern, optimization-based, and inlier-like backdoor attacks. Therefore, going beyond single-view aggregation, this thesis presents a multi-view end-to-end Byzantine-resilient blockchain-enabled FL framework. Our framework couples permissioned multi-channel blockchain orchestration with decentralized off-chain model storage. In addition, the proposed Byzantine-resilient blockchain-based FL framework introduces a new multi-view defense mechanism — Curvature-Statistics Cross-Validation Defense (CSCVD). The proposed blockchain integrated CSCVD defense mechanism is used to detect and suppress backdoored updates while preserving clean-task performance of the proposed framework. The proposed framework is evaluated under non-IID federated settings with a standard multi-client training protocol. Experimental evaluations are conducted on three image classification benchmarks (CIFAR-10, CIFAR-100, and EMNIST) and tested against three increasingly strong backdoor threats, namely, Blackbox Poisoning, Projected Gradient Descent (PGD) backdoor optimization under update constraints, and PGD with Model Replacement (PGD+MR). Our results indicate that optimization-based stealth and model-replacement scaling cannot simultaneously bypass our proposed CSCVD’s multi-view constraints. For example, on CIFAR-10 under PGD+MR attack, backdoor success is reduced to 2.76% while maintaining 88.44% validation accuracy by CSCVD. Additional evaluations highlight how the robustness behavior of CSCVD is retained under coordinated multi-attacker collusion even when attackers synchronize their malicious updates. Overall, this work contributes a new deployable blueprint for Byzantine-resilient Federated Learning by combining blockchain-based accountability with a multi-view internal consistency defense.

 

Presenter: Muhammad Golam Rohman Shafi (Std No. 0424052105)

Venue: Graduate Seminar Room