Abstract
In resource-constrained Internet of Things (IoT) scenarios, implementing robust and accurate deep learning inference is problematic due to device failures, limited computing power, and privacy concerns. We present a resilient, completely edge-based distributed convolutional neural network (CNN) architecture that eliminates cloud dependencies while enabling accurate and fault-tolerant inference. At its core is a lightweight Modifier Module deployed at the edge, which synthesizes predictions for failing devices by pooling peer CNN outputs and weights. This dynamic mechanism is trained via a novel fail-simulation technique, allowing it to mimic missing outputs in real-time without model duplication or cloud fallback. We assess our methodology using MNIST and CIFAR-10 datasets under both homogeneous and heterogeneous data partitions, with up to five simultaneous device failures. The system displays up to 1.5% absolute accuracy improvement, 30% error rate reduction, and stable operation even with over 80% device dropout, exceeding ensemble, dropout, and federated baselines. Our strategy combines significant statistical significance, low resource utilization (~ 15 KB per model), and real-time responsiveness, making it well-suited for safety-critical IoT installations where cloud access is infeasible.