Search papers, labs, and topics across Lattice.
This paper introduces SABLE, a novel backdoor attack in federated learning (FL) that uses semantically meaningful, in-distribution triggers, such as adding sunglasses to faces or altering traffic sign colors. SABLE optimizes a malicious objective with feature separation and parameter regularization to maintain stealth while achieving high attack success rates. Experiments on CelebA and GTSRB datasets demonstrate that SABLE achieves high targeted attack success rates across various FL aggregation rules, highlighting the vulnerability of FL to semantics-aligned backdoors.
FL systems are far more vulnerable to backdoor attacks using realistic, semantically-aligned triggers (like sunglasses) than previously thought based on simple corner patches.
Backdoor attacks on federated learning (FL) are most often evaluated with synthetic corner patches or out-of-distribution (OOD) patterns that are unlikely to arise in practice. In this paper, we revisit the backdoor threat to standard FL (a single global model) under a more realistic setting where triggers must be semantically meaningful, in-distribution, and visually plausible. We propose SABLE, a Semantics-Aware Backdoor for LEarning in federated settings, which constructs natural, content-consistent triggers (e.g., semantic attribute changes such as sunglasses) and optimizes an aggregation-aware malicious objective with feature separation and parameter regularization to keep attacker updates close to benign ones. We instantiate SABLE on CelebA hair-color classification and the German Traffic Sign Recognition Benchmark (GTSRB), poisoning only a small, interpretable subset of each malicious client's local data while otherwise following the standard FL protocol. Across heterogeneous client partitions and multiple aggregation rules (FedAvg, Trimmed Mean, MultiKrum, and FLAME), our semantics-driven triggers achieve high targeted attack success rates while preserving benign test accuracy. These results show that semantics-aligned backdoors remain a potent and practical threat in federated learning, and that robustness claims based solely on synthetic patch triggers can be overly optimistic.
