Trust-free Personalized Decentralized Learning

TL;DR

Trust-free Personalized Decentralized Federated Learning (TPFed) proposes a blockchain-backed bulletin board to enable global neighbor discovery via Locality-Sensitive Hashing (LSH) and Peer Ranking (PR) in a fully decentralized federated setting. It introduces an all-in-one knowledge distillation protocol that transfers knowledge, evaluates model quality, and verifies similarity using a public reference dataset, while protecting local data. The framework incorporates robust countermeasures against LSH forgery, collusion, and poisoning, including a commit-then-reveal scheme and a consistency check based on $D_{KL}$ divergences. Empirical results on MNIST, A-ECG, S-EEG, and CIFAR-100 show that TPFed achieves superior accuracy and resilience compared to centralized, semi-decentralized, and decentralized baselines, with a scalable variant TPFed-idx offering improved efficiency. This work advances secure, globally personalized collaboration in open, trust-averse environments with potential applications in privacy-sensitive domains such as e-health and decentralized finance, while noting areas for future work in heterogeneous architectures and incentive mechanisms.

Abstract

Personalized collaborative learning in federated settings faces a critical trade-off between customization and participant trust. Existing approaches typically rely on centralized coordinators or trusted peer groups, limiting their applicability in open, trust-averse environments. While recent decentralized methods explore anonymous knowledge sharing, they often lack global scalability and robust mechanisms against malicious peers. To bridge this gap, we propose TPFed, a \textit{Trust-free Personalized Decentralized Federated Learning} framework. TPFed replaces central aggregators with a blockchain-based bulletin board, enabling participants to dynamically select global communication partners based on Locality-Sensitive Hashing (LSH) and peer ranking. Crucially, we introduce an ``all-in-one'' knowledge distillation protocol that simultaneously handles knowledge transfer, model quality evaluation, and similarity verification via a public reference dataset. This design ensures secure, globally personalized collaboration without exposing local models or data. Extensive experiments demonstrate that TPFed significantly outperforms traditional federated baselines in both learning accuracy and system robustness against adversarial attacks.

Figures (4)

• Figure 1: Overview of the TPFed framework. The central coordinator is replaced by a blockchain-based bulletin board, which filters neighbor candidates based on inter-client similarity (via LSH) and model quality (via PR). Clients then employ all-in-one logits to simultaneously perform knowledge transfer, model quality evaluation, and inter-client similarity measurement, where the latter also verifies the LSH and PR claims. ${}^{*}$In practice, clients publish commitments $C$ in the bulletin instead of the original neighbor ranking PR; however, PR is depicted in the figure to simplify the illustration of data flow.
• Figure 2: Client learns from neighbors via collecting logits $\hat{Y}^{\text{web}}$.
• Figure 3: One LSH Cheating Attack Start from $t=50$.
• Figure 4: Poison attacks with varying amounts of dishonest clients. Attacks start from the 50th iteration. Blue curve: TPFed; red curve: ProxyFL. Shaded areas denote the standard deviation across honest clients.