Modifier guided resilient CNN inference enables fault-tolerant edge collaboration for IoT

Area of Science:

• Artificial Intelligence
• Computer Science
• Internet of Things

Background:

• Resource-constrained Internet of Things (IoT) environments face challenges in deep learning inference due to device failures, limited computing power, and privacy concerns.
• Cloud-dependent solutions are often unsuitable for IoT due to these limitations and the need for real-time, localized processing.

Purpose of the Study:

• To develop a resilient, completely edge-based distributed convolutional neural network (CNN) architecture for accurate and fault-tolerant deep learning inference in IoT.
• To eliminate cloud dependencies while maintaining high performance and robustness against device failures.

Main Methods:

• A lightweight Modifier Module was developed and deployed at the edge to synthesize predictions for failing devices by pooling outputs and weights from peer CNNs.
• A novel fail-simulation technique was employed to train the Modifier Module for real-time mimicry of missing outputs without model duplication or cloud fallback.
• The methodology was evaluated using MNIST and CIFAR-10 datasets under various data partitioning scenarios, simulating up to five simultaneous device failures.

Main Results:

• The proposed system demonstrated up to 1.5% absolute accuracy improvement and a 30% error rate reduction.
• The system maintained stable operation with over 80% device dropout, outperforming ensemble, dropout, and federated learning baselines.
• The solution exhibited low resource utilization (approximately 15 KB per model) and real-time responsiveness.

Conclusions:

• The developed edge-based distributed CNN architecture offers a robust, accurate, and fault-tolerant solution for deep learning inference in resource-constrained IoT scenarios.
• The system's ability to operate without cloud dependencies, combined with its low resource utilization and resilience to device failures, makes it ideal for safety-critical IoT applications.