Researchers identify vulnerability in neural networks that increases their energy consumption

A new type of hacker attack could increase the energy consumption of AI systems. Researchers from the Maryland Center for Cybersecurity reached this conclusion, according to Technology Review.

The threat operates similarly to a DoS attack on the internet, aiming to “clog” the neural network and force it to use more computing resources and slow its “thinking” process, according to experts.

The vulnerability concerns input-adaptive architectures with multiple outputs that allocate tasks depending on the difficulty of solving them and optimize the use of computing resources. By adding a small amount of noise to inputs, researchers made the network perceive them as more complex and increase the volume of computations.

Experimentally, researchers found that an attacker can maximally overload the neural network if he has full information about it. When researchers assumed that the hacker had no information, they were able to slow the algorithm and increase energy consumption by 20–80%.

According to the researchers, this is because attacks transfer well across different types of neural networks.

“Developing an attack for one image-classification system is enough to disrupt the operation of many others”, said a graduate student and co-author Yiğitcan Kaya [Yiğitcan Kaya].

The group noted that the vulnerability is theoretical. Adaptive-input architectures are not yet used in real-world applications. However the situation may change in the near future under industry pressure to deploy lighter neural networks, for example in the Internet of Things.

In early May 2021, Microsoft released an open-source tool that helps developers test the security of AI systems.

In April, Microsoft integrated threat-detection technology from Intel for detecting hidden miners into the Defender for Endpoint security platform for enterprises.

At the end of 2020, the largest hacker attack in U.S. history occurred via vulnerabilities in SolarWinds software. It affected, among others, the State Department, several ministries, and partly the Pentagon.

Subscribe to ForkLog’s news on Telegram: ForkLog AI — all the AI news!