Self-healing neural networks via modular patch layers for diverse structural and adversarial damages

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep neural networks (DNNs) are increasingly used in safety-critical systems, raising concerns about their reliability and continuous operation.
• Internal flaws or adversarial attacks can degrade DNN performance, impacting system dependability.
• Maintaining the operational integrity of deployed DNNs presents a significant challenge.

Purpose of the Study:

• To present a novel Self-Healing Neural Network (SHNN) model for post-deployment recovery of DNNs.
• To develop a mechanism for automatic identification, localization, and repair of damaged network components.
• To enhance the resilience and fault tolerance of AI systems in high-reliability environments.

Main Methods:

• Implemented a Self-Healing Neural Network (SHNN) model utilizing lightweight, modular patch layers.
• Employed activation discrepancy analysis at each layer to identify and isolate faults.
• Performed selective retraining of afflicted layers, avoiding full retraining or global fine-tuning.
• Integrated a live Streamlit interface for demonstration and interaction.

Main Results:

• SHNN restored performance after structural damage to near-baseline accuracy levels (e.g., 97.33% vs. 97.76%).
• The model recovered 74-84% of accuracy lost due to FGSM and PGD adversarial attacks.
• Post-healing activation deviation scores significantly decreased, confirming effective fault localization.
• Demonstrated SHNN's modular, interpretable, and dataset-adaptive nature.

Conclusions:

• The SHNN framework provides an effective self-repair mechanism for deep neural networks.
• SHNN enhances the reliability and fault tolerance of AI systems, crucial for safety-critical applications.
• The modular and adaptive design allows for broad applicability in high-reliability environments.