This study explores the comparative strengths of distributed learning models, Federated Learning (FL), Blind Federated Learning (BFL), Blended Blind Federated Learning (BBFL), Split Learning (SL), and Decentralized Learning (DL), by evaluating their performance metrics, convergence rates, and security features. Distributed learning models aim to leverage data from multiple clients while maintaining privacy. Each model uses different architectural and security mechanisms to achieve this, resulting in unique strengths and limitations in terms of scalability, resilience to attacks, and data integrity. By comparing convergence speed, accuracy, and resilience to specific threats, this paper provides insights into each model’s effectiveness for privacy-sensitive applications, particularly in handling model inversion, gradient leakage, and data poisoning. Results suggest that while DL exhibits superior performance and security, BBFL and SL models provide competitive alternatives for structured environments with moderate security needs. This analysis offers a framework to guide distributed learning model selection based on application requirements and security priorities.
Post-training quantization (PTQ) can make medical image segmentation models smaller and faster without retraining, but its effect on accuracy is strongly dependent on the task, calibration data, and network architecture. We present Sensitivity-Guided Post-Training Quantization (SGQ), a deployment-oriented INT8 framework that combines depth-balanced calibration, ONNX/QDQ static quantization, and layer-wise sensitivity analysis. SGQ identifies layers where INT8 conversion is fragile, supports selective INT8/FP32 node exclusion, and turns the sensitivity map into practical deployment guidance. Across three Medical Segmentation Decathlon tasks and two U-Net-style architectures, calibration choice is the main determinant of PTQ stability. Percentile calibration is the only tested method that avoids catastrophic failure across all six task-model combinations, while MSE calibration fails severely in three combinations, with Dice drops of 19.4, 38.8, and 67.1 percentage points. CT windowing experiments further show that the best calibration strategy depends on the target preprocessing pipeline. Under wide HU windowing, StackLite25D is adequately served by MinMax calibration, whereas UNet2D still benefits from Percentile calibration. A ten-model validation on Hepatic Vessel segmentation shows that large first-layer sensitivity is task-activated rather than architecture-intrinsic. It appears on sparse thin-structure targets and is absent on compact organ tasks under the same SGQ threshold. Measured benchmarks on four mobile devices show FP32-to-INT8-SGQ speedups of 1.51 to 2.84×, while paired macOS CPU measurements report latency, throughput, and memory footprint. Overall, SGQ provides a practical workflow for medical INT8 deployment: choose calibration first, profile sparse targets, and reserve FP32 only for layers whose sensitivity justifies protection.