Hybrid GNN-LSTM defense with differential privacy and secure multi-party computation for edge-optimized neuromorphic...

Area of Science:

• Artificial Intelligence
• Computer Vision
• Cybersecurity

Background:

• Neuromorphic computing, utilizing spiking neural networks (SNNs) and event cameras, offers energy-efficient perception for autonomous vehicles (AVs).
• These systems are vulnerable to adversarial attacks, fault injections, and data poisoning, posing significant security risks.
• Existing defenses are often insufficient for resource-constrained edge environments.

Purpose of the Study:

• To develop a robust and efficient security solution for neuromorphic autonomous systems.
• To address the susceptibility of these systems to various adversarial threats.
• To ensure privacy-conscious and resource-efficient security for next-generation AVs.

Main Methods:

• Proposed a Hybrid Graph Neural Network-Long Short-Term Memory (GNN-LSTM) model for attack detection.
• Integrated Differential Privacy (DP) and Secure Multi-Party Computation (SMPC) for privacy and threat reduction.
• Applied quantization and pruning techniques to optimize the framework for edge deployment.

Main Results:

• Achieved 94.3% accuracy on KITTI multimodal experiments, reducing attack success rates by 30%.
• Reached 92.4% accuracy on neuromorphic N-Caltech101 experiments with a 27% drop in attack success.
• Demonstrated the effectiveness of the proposed solution against trained adversarial attacks.

Conclusions:

• The developed security solution provides substantial, privacy-conscious, and resource-efficient protection for neuromorphic AVs.
• The Hybrid GNN-LSTM model, combined with DP and SMPC, effectively mitigates adversarial threats.
• Optimized framework ensures feasibility for edge deployment in autonomous systems.