GPU Memory Exploitation for Fun and Profit

As modern applications increasingly rely on GPUs to accelerate the computation, it has become very critical to study and understand the security implications of GPUs. In this work, we conduct a thorough examination of buffer overflows on modern GPUs. Specifically, we demonstrate that, due to GPU's unique memory system, GPU programs suffer from different and more complex buffer overflow vulnerabilities compared to CPU programs, contradicting the conclusions of prior studies. In addition, despite the critical role GPUs play in modern computing, GPU systems are missing essential memory protection mechanisms. Consequently, when buffer overflow vulnerabilities are exploited by an attacker, they can lead to both code injection attacks and code reuse attacks, including return-oriented programming (ROP). Our results show that these attacks pose a significant security risk to modern GPU applications.