Single-Bit Flips Can Sabotage AI Models, Researchers Find

Researchers at George Mason University have uncovered a critical vulnerability in DEEP learning models. By flipping just one bit in memory—changing a 1 to a 0 or vice versa—attackers can subtly alter a model's behavior without retraining, rewriting code, or significantly degrading accuracy. The sabotage reduces model performance by less than 0.1%, making it nearly undetectable.

The attack exploits the 'Rowhammer' hardware vulnerability, where repeated memory access causes adjacent bits to flip unintentionally. Hackers can execute this by running malicious code on the same machine as the AI, whether through malware, compromised cloud accounts, or other vectors. The targeted bit, once modified in RAM, introduces a hidden backdoor.

This flaw poses risks for sensitive applications like self-driving cars and medical AI, where even minor deviations could have catastrophic consequences. The research highlights the fragility of AI systems under adversarial conditions—a growing concern as these models permeate critical infrastructure.