אירועים והרצאות בפקולטה למדעי המחשב ע"ש הנרי ומרילין טאוב

The extraction of neural networks poses a significant challenge to the security and intellectual property of AI models, enabling adversaries to recreate proprietary architectures, breach confidentiality, and exploit model functionality. In this seminar talk, I will introduce a novel attack that reconstructs both the structure and exact parameters of black-box convolutional neural networks (CNNs), using only query-based access. This technique is the first to recover the precise weight values and architecture of black-box CNNs. This method allows the extraction of common CNN models, including LeNet-5, AlexNet, and various VGG and ResNet architectures. I will outline the theoretical foundations of the attack and demonstrate its effectiveness through extractions of multiple architectures. This work highlights the real-world feasibility of model extraction and its broader implications for AI security.