همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Optimized Reconstructive Pruning for Trojan Detection in Neural Networks with RL-GNN

نویسندگان :

Mohsen Askari¹ Shahrooz Janbaz² Abolfazl Rahmani³

1- Malek Ashtar University of Technology 2- Malek Ashtar University of Technology 3- Malek Ashtar University of Technology

کلمات کلیدی :

Trojan detection،deep neural networks،reconstructive pruning،reinforcement learning،graph neural networks،AI security

چکیده :

Trojan attacks, embedding malicious modifications in deep neural networks (DNNs), pose significant threats to their reliability in safety-critical applications, such as autonomous driving and medical diagnostics. These stealthy backdoors trigger erroneous outputs on specific inputs while maintaining normal behavior otherwise, challenging traditional validation methods. This study proposes an Optimized Reconstructive Pruning (ORP) framework to enhance Trojan detection across diverse DNN architectures, including convolutional networks and transformers. ORP integrates reconstructive pruning, which mitigates accuracy degradation post-pruning, with a novel hybrid Reinforcement Learning-Graph Neural Network (RL-GNN) optimizer. This optimizer dynamically navigates complex pruning parameter spaces, improving scalability and efficiency. Evaluated on TrojAI benchmarks comprising over 2,500 models, ORP achieves a classification accuracy of 92.3%, surpassing baseline pruning methods by 16%, while reducing execution time by 36%. The framework demonstrates robustness in low-resource settings and noisy environments, effectively amplifying Trojan-induced asymmetries. These advancements enhance the detection of sophisticated backdoors, ensuring reliable DNN deployment. Future research may explore ORP’s integration into federated learning and its applicability to emerging adaptive threats, strengthening AI security in real-world systems.