Researchers from the University of Toronto have demonstrated that Rohammer attacks can be conducted against GPUs, specifically Nvidia GDDR6 memory, affecting machine learning models. The findings highlight the potential security risks to AI systems and the importance of system-level ECC mitigation. #GPUHammer #Rohammer #NvidiaA6000 #DeepNeuralNetworks
Keypoints
- The University of Toronto researchers proved Rohammer attacks are feasible on GPUs, specifically Nvidia’s architecture.
- GPUHammer can cause bit flips in GDDR6 memory, significantly degrading machine learning model accuracy.
- Such attacks could lead to data corruption, privilege escalation, or breaking memory isolation in virtualized environments.
- Nvidia confirmed the vulnerability and recommends system-level ECC to prevent Rohammer attacks, though it may impact performance.
- Testing against other GPUs remains challenging due to soldered memory modules, increasing research costs.
Read More: https://www.securityweek.com/rowhammer-attack-demonstrated-against-nvidia-gpu/