DivQAT: Robust Quantized CNNs Against Extraction Attacks

This paper addresses the vulnerability of quantized Convolutional Neural Networks (CNNs) to model extraction attacks, a critical issue for intellectual property protection. It introduces DivQAT, a novel training algorithm that integrates defense mechanisms directly into the quantization process. This is a significant contribution because it moves beyond post-training defenses, which are often computationally expensive and less effective, especially for resource-constrained devices. The paper's focus on quantized models is also important, as they are increasingly used in edge devices where security is paramount. The claim of improved effectiveness when combined with other defense mechanisms further strengthens the paper's impact.