"SADC: State-Aware and Dynamically Calibrated Security Control for Non-Stationary Backdoor Attacks"
Abstract
"Backdoor defense in federated learning (FL) is a security-control process that identifies, suppresses, and isolates malicious model updates while preserving privacy and model util ity. Its central difficulty is that the attack intensity, participation pattern, and malicious population may vary over time, whereas existing defenses often rely on fixed thresholds and predefined rules. Such a mismatch makes it difficult to maintain a stable balance among robustness, privacy protection, communication efficiency, and model utility in non-stationary adversarial envi ronments. To address this problem, we develop SADC, a state aware and dynamically calibrated security-control framework for FL backdoor defense. First, SADC adjusts defense parameters according to calibration feedback and blind-state risk statistics so that the defense strength can follow changes in training state and attack risk. Second, SADC combines adaptive screening, robust aggregation, and energy-compensated reconstruction to suppress malicious updates while preserving useful update information. Finally, SADC incorporates differential privacy constraints into a privacy-preserving compressed upload process for honest clients. Experiments against model replacement, dynamic backdoor, and distributed backdoor attacks on MNIST, Fashion-MNIST, and CIFAR-10 show that SADC maintains ASR below 3% across all nine evaluated dataset\u2013attack combinations while preserving competitive task accuracy. In the representative…
Read the paper
Related studies
- Active Learning during Federated Learning for Object Detection
- Fairness-aware federated learning
- Decentralized Federated Learning
- Building a Personally Identifiable Information Recognizer in a Privacy Preserved Manner Using Automated Annotation and Federated Learning
- Preservation of Patient Level Privacy: Federated Classification and Calibration Models
- Efficient explainable federated learning
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.