Deep neural networks (DNNs) have been a very active field of research for eight years now, and for the last five we’ve seen a steady stream of adversarial examples – inputs that will bamboozle a DNN so that it thinks a 30mph speed limit sign is a 60 instead, and even magic spectacles to make a DNN get the wearer’s gender wrong.
So far, these attacks have targeted the integrity or confidentiality of machine-learning systems. Can we do anything about availability?
Sponge Examples: Energy-Latency Attacks on Neural Networks shows how to find adversarial examples that cause a DNN to burn more energy, take more time, or both. They affect a wide range of DNN applications, from image recognition to natural language processing (NLP). Adversaries might use these examples for all sorts of mischief – from draining mobile phone batteries, though degrading the machine-vision systems on which self-driving cars rely, to jamming cognitive radar.
So far, our most spectacular results are against NLP systems. By feeding them confusing inputs we can slow them down over 100 times. There are already examples in the real world where people pause or stumble when asked hard questions but we now have a dependable method for generating such examples automatically and at scale. We can also neutralize the performance improvements of accelerators for computer vision tasks, and make them operate on their worst case performance.
One implication is that engineers designing real-time systems that use machine learning will have to pay more attention to worst-case behaviour; another is that when custom chips used to accelerate neural network computations use optimisations that increase the gap between worst-case and average-case outcomes, you’d better pay even more attention.