New cyber-physical attack manipulates data center power consumption

As reported by Cyber Insider, a novel cyber-physical attack named Bit2Watt has been demonstrated, capable of manipulating a data center's power consumption through carefully crafted GPU workloads. This emerging security concern highlights the potential risks as AI data centers become increasingly integrated with modern electrical infrastructure, particularly those powered by renewable energy sources.

Researchers at Zhejiang University have developed Bit2Watt, a proof-of-concept attack that abuses rented GPUs by running specialized workloads designed to rapidly fluctuate between high and low computational intensity. These fluctuations create high-frequency power draw changes that can propagate into the local electrical grid. The attack, dubbed "Bit2Watt" for its software-to-power influence, also has a potential feedback effect called "Watt2Bit," where degraded power quality could disrupt computing equipment or lead to information leakage. Laboratory tests and simulations showed that synchronized attacks on a large scale could lead to significant harmonic distortion and system instability.

The researchers note that detecting such attacks is challenging as the workloads are legitimate and standard monitoring tools may not capture the high-frequency power signatures. They suggest joint defenses involving workload scheduling, enhanced power monitoring, and better coordination between data centers and power system operators.

Source: Cyber Insider