.png){kind=icon}
.png){kind=icon}
Return
Privacy-Preserving Federated Multimodal Agriproduct Anomaly Detection in AIoT via Modality-Under-Optimized Knowledge Distillation
.png){kind=icon}
PRE
AI
J
M
C
Y
L
Q
DOI：10.1109/jiot.2026.3703645.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
As a key application of the Agricultural Internet of Things (AIoT), multimodal agriproduct anomaly detection faces severe privacy and security challenges. Existing federated learning (FL) methods struggle to capture fine-grained cross-modal correlations and to address the modality under-optimization problem, thereby limiting both detection accuracy and privacy levels. To this end, this article proposes a privacy-preserving federated multimodal agriproduct anomaly detection scheme in AIoT based on modality-under-optimized knowledge distillation (PAMAD), achieving high-utility anomaly detection with enhanced privacy protection. Specifically, we develop a hierarchical privacy protection method for multimodal fine-grained alignment fusion based on meta-learning (HPPMF), which effectively captures cross-modal semantic correlations and protects the privacy of fused features. In addition, we propose a multitask pretraining algorithm based on modality-under-optimized knowledge distillation (MTLMKD) to alleviate modal imbalance. We further design a pretraining dynamic protection algorithm based on adaptive gradient quantization (PDAGC) to ensure model security. Subsequently, a multimodal agriproduct anomaly detection method with a self-supervised denoising encoder (MAPADSE) is introduced to improve detection accuracy under noisy conditions. Rigorous security analysis demonstrates that the PAMAD scheme satisfies differential privacy (DP). The experimental results show that, compared with existing state-of-the-art methods, our PAMAD scheme improves AUROC and accuracy by 7.56% and 8.71%, respectively, achieving a desirable balance between privacy protection and anomaly detection accuracy in AIoT services.
Keywords:
Agriproduct anomaly detection
differential privacy (DP)
federated learning (FL)
knowledge distillation
Journal
IF:
8.9
Papers:
1.4W
Citations:
7.8W
Organization
Cited Papers