ROWHAMMER ATTACK CAN BACKDOOR AI MODELS WITH ONE DEVASTATING BIT FLIP

Summary from the TLDR InfoSec newsletter

ONEFLIP is a Rowhammer-based attack that can insert backdoors into AI models by flipping just a single bit in DRAM memory during inference. The attack targets vulnerable weights in the final classification layer of neural networks, allowing attackers to craft imperceptible trigger patterns that cause catastrophic misclassification (such as making self-driving cars misidentify stop signs as speed limit signs) while maintaining normal model accuracy for other inputs. It requires access to a white-box model and co-location on the target machine with DDR3/DDR4 memory, making it particularly concerning for multi-tenant cloud platforms and AI inference servers.