FedRACE: A Hierarchical and Statistical Framework for Robust Federated Learning

Integrating large pre-trained models into federated learning (FL) can significantly improve generalization and convergence efficiency. A widely adopted strategy freezes the pre-trained backbone and fine-tunes a lightweight task head, thereby reducing computational and communication costs. However, this partial fine-tuning paradigm introduces new security risks, making the system vulnerable to poisoned updates and backdoor attacks. To address these challenges, we propose FEDRACE, a unified framework for robust FL with partially frozen models. FEDRACE comprises two core components: HStat-Net, a hierarchical network that refines frozen features into compact, linearly separable representations; and DevGuard, a serverside mechanism that detects malicious clients by evaluating statistical deviance in class-level predictions modeling generalized linear models (GLMs). DevGuard further incorporates adaptive thresholding based on theoretical misclassification bounds and employs randomized majority voting to enhance detection reliability. We implement FEDRACE on the FedScale platform and evaluate it on CIFAR-100, Food-101, and Tiny ImageNet under diverse attack scenarios. FEDRACE achieves a true positive rate of up to 99.3% with a false positive rate below 1.2%, while preserving model accuracy and improving generalization.

Existing defenses such as Trimmed-Mean [19], Multi-Krum [20], and reputation-based methods like FLShield [21] and FLAIR [22] often rely on gradient statistics or fixed heuristics. While these methods are effective in some settings, they struggle to detect subtle semantic manipulations, especially when only the head is trainable [23,24]. These limitations raise a key research question: How can we integrate large pre-trained models into FL while enabling reliable and adaptive detection of malicious clients?

To answer this question, we propose FEDRACE, a unified framework for Federated Representationbased Adaptive Client Evaluation. FEDRACE combines hierarchical representation learning with statistical client evaluation to improve FL robustness. It consists of two main components: (1) HStat-Net, a Hierarchical Statistical Network that transforms fixed features into compact and linearly separable representations using a triplet loss, and (2) DevGuard, a server-side evaluation mechanism that uses a generalized linear model (GLM) to identify clients with abnormal semantic behavior through deviance analysis.