New Hack Forces Energy-Efficient AI Systems to Waste Power
Researchers reveal a novel adversarial attack that forces efficient AI models to consume significantly more energy by tricking them into perceiving simple inputs as highly complex.
Artificial intelligence systems consume massive amounts of power, and a newly discovered vulnerability allows hackers to force these models to waste even more energy. Similar to a denial-of-service attack that clogs a network, this new method targets deep neural networks and compels them to tie up unnecessary computational resources, which ultimately slows down their "thinking" process.
The specific target of this attack is a highly efficient category of neural networks designed to reduce AI's carbon footprint. These input-adaptive multi-exit architectures save power by splitting tasks based on difficulty, spending minimal computation on easy inputs like a clear, well-lit photo and reserving deeper processing only for complex images.
Researchers from the Maryland Cybersecurity Center discover that adding small amounts of noise to the network's input tricks the system into perceiving easy tasks as incredibly difficult, which forces it to max out its energy draw. Even when attackers possess limited knowledge about the AI model, they successfully increase its energy consumption by 20% to 80%, posing a significant threat to devices like smartphones and smart speakers.