Privacy Preserving Image Encryption with Optimal Deep Transfer Learning Based Accident Severity Classification Model.

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Effective accident management relies on timely data analysis, with images being crucial evidence.
• Image data in accident management systems face privacy risks and transmission challenges.
• Automated accident severity classification using deep learning is essential for efficient emergency response.

Purpose of the Study:

• To propose a novel Privacy Preserving Image Encryption with Optimal Deep-Learning-based Accident Severity Classification (PPIE-ODLASC) method.
• To ensure secure transmission and storage of accident images through encryption.
• To accurately classify accident severity using advanced deep learning techniques.

Main Methods:

• Image encryption using multi-key homomorphic encryption (MKHE) with lion swarm optimization (LSO) for key generation.
• Accident severity classification employing YOLO-v5 for region of interest identification, Xception for feature extraction, and bidirectional gated recurrent unit (BiGRU) for classification.
• Hyperparameter tuning using Bayesian optimization (BO) for enhanced model performance.

Main Results:

• The PPIE-ODLASC method demonstrated robust performance in securely encrypting and classifying accident images.
• Experimental validation showed significant improvements, with a 57.68 dB enhancement over existing models.
• The integrated approach effectively balances privacy preservation with accurate severity assessment.

Conclusions:

• The PPIE-ODLASC method offers a secure and efficient solution for accident image management and severity classification.
• This approach enhances the capabilities of emergency and traffic control systems through advanced AI and encryption.
• The findings highlight the potential of combining privacy-preserving techniques with deep learning for critical infrastructure applications.