arXiv:2501.12123v2 Announce Type: replace-cross Abstract: Federated Learning (FL) enables clients to collaboratively train a global model using their local datasets while reinforcing data privacy, but it is prone to poisoning attacks. Existing defense mechanisms assume that clients' data are independent and identically distributed (IID), making them ineffective in real-world applications where data are non-IID. This paper presents FL-CLEANER, the first defense capable of filtering both byzantine and backdoor attackers' model updates in a non-IID FL environment. The originality of FL-CLEANER is twofold. First, it relies on a client confidence score derived from the reconstruction errors of each client's model activation maps for a given trigger set, with reconstruction errors obtained by means of a Conditional Variational Autoencoder trained according to a novel server-side strategy. Second, it uses an original ad-hoc trust propagation algorithm we propose. Based on previous client scores, it allows building a cluster of benign clients while flagging potential attackers. Experimental results on the datasets MNIST and FashionMNIST demonstrate the efficiency of FL-CLEANER against Byzantine attackers as well as to some state-of-the-art backdoors in non-IID scenarios; it achieves a close-to-zero (<1%) benign client misclassification rate, even in the absence of an attack, and achieves strong performance compared to state of the art defenses.